2 * @brief memory-mapped database library
4 * A Btree-based database management library modeled loosely on the
5 * BerkeleyDB API, but much simplified.
8 * Copyright 2011-2013 Howard Chu, Symas Corp.
11 * Redistribution and use in source and binary forms, with or without
12 * modification, are permitted only as authorized by the OpenLDAP
15 * A copy of this license is available in the file LICENSE in the
16 * top-level directory of the distribution or, alternatively, at
17 * <http://www.OpenLDAP.org/license.html>.
19 * This code is derived from btree.c written by Martin Hedenfalk.
21 * Copyright (c) 2009, 2010 Martin Hedenfalk <martin@bzero.se>
23 * Permission to use, copy, modify, and distribute this software for any
24 * purpose with or without fee is hereby granted, provided that the above
25 * copyright notice and this permission notice appear in all copies.
27 * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
28 * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
29 * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
30 * ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
31 * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
32 * ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
33 * OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
38 #include <sys/types.h>
40 #include <sys/param.h>
46 #ifdef HAVE_SYS_FILE_H
63 #if !(defined(BYTE_ORDER) || defined(__BYTE_ORDER))
64 #include <netinet/in.h>
65 #include <resolv.h> /* defines BYTE_ORDER on HPUX and Solaris */
68 #if defined(__APPLE__) || defined (BSD)
69 # define MDB_USE_POSIX_SEM 1
70 # define MDB_FDATASYNC fsync
71 #elif defined(ANDROID)
72 # define MDB_FDATASYNC fsync
77 #ifdef MDB_USE_POSIX_SEM
78 #include <semaphore.h>
83 #include <valgrind/memcheck.h>
84 #define VGMEMP_CREATE(h,r,z) VALGRIND_CREATE_MEMPOOL(h,r,z)
85 #define VGMEMP_ALLOC(h,a,s) VALGRIND_MEMPOOL_ALLOC(h,a,s)
86 #define VGMEMP_FREE(h,a) VALGRIND_MEMPOOL_FREE(h,a)
87 #define VGMEMP_DESTROY(h) VALGRIND_DESTROY_MEMPOOL(h)
88 #define VGMEMP_DEFINED(a,s) VALGRIND_MAKE_MEM_DEFINED(a,s)
90 #define VGMEMP_CREATE(h,r,z)
91 #define VGMEMP_ALLOC(h,a,s)
92 #define VGMEMP_FREE(h,a)
93 #define VGMEMP_DESTROY(h)
94 #define VGMEMP_DEFINED(a,s)
98 # if (defined(_LITTLE_ENDIAN) || defined(_BIG_ENDIAN)) && !(defined(_LITTLE_ENDIAN) && defined(_BIG_ENDIAN))
99 /* Solaris just defines one or the other */
100 # define LITTLE_ENDIAN 1234
101 # define BIG_ENDIAN 4321
102 # ifdef _LITTLE_ENDIAN
103 # define BYTE_ORDER LITTLE_ENDIAN
105 # define BYTE_ORDER BIG_ENDIAN
108 # define BYTE_ORDER __BYTE_ORDER
112 #ifndef LITTLE_ENDIAN
113 #define LITTLE_ENDIAN __LITTLE_ENDIAN
116 #define BIG_ENDIAN __BIG_ENDIAN
119 #if defined(__i386) || defined(__x86_64) || defined(_M_IX86)
120 #define MISALIGNED_OK 1
126 #if (BYTE_ORDER == LITTLE_ENDIAN) == (BYTE_ORDER == BIG_ENDIAN)
127 # error "Unknown or unsupported endianness (BYTE_ORDER)"
128 #elif (-6 & 5) || CHAR_BIT != 8 || UINT_MAX < 0xffffffff || ULONG_MAX % 0xFFFF
129 # error "Two's complement, reasonably sized integer types, please"
132 /** @defgroup internal MDB Internals
135 /** @defgroup compat Windows Compatibility Macros
136 * A bunch of macros to minimize the amount of platform-specific ifdefs
137 * needed throughout the rest of the code. When the features this library
138 * needs are similar enough to POSIX to be hidden in a one-or-two line
139 * replacement, this macro approach is used.
143 #define pthread_t DWORD
144 #define pthread_mutex_t HANDLE
145 #define pthread_key_t DWORD
146 #define pthread_self() GetCurrentThreadId()
147 #define pthread_key_create(x,y) \
148 ((*(x) = TlsAlloc()) == TLS_OUT_OF_INDEXES ? ErrCode() : 0)
149 #define pthread_key_delete(x) TlsFree(x)
150 #define pthread_getspecific(x) TlsGetValue(x)
151 #define pthread_setspecific(x,y) (TlsSetValue(x,y) ? 0 : ErrCode())
152 #define pthread_mutex_unlock(x) ReleaseMutex(x)
153 #define pthread_mutex_lock(x) WaitForSingleObject(x, INFINITE)
154 #define LOCK_MUTEX_R(env) pthread_mutex_lock((env)->me_rmutex)
155 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock((env)->me_rmutex)
156 #define LOCK_MUTEX_W(env) pthread_mutex_lock((env)->me_wmutex)
157 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock((env)->me_wmutex)
158 #define getpid() GetCurrentProcessId()
159 #define MDB_FDATASYNC(fd) (!FlushFileBuffers(fd))
160 #define MDB_MSYNC(addr,len,flags) (!FlushViewOfFile(addr,len))
161 #define ErrCode() GetLastError()
162 #define GET_PAGESIZE(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;}
163 #define close(fd) (CloseHandle(fd) ? 0 : -1)
164 #define munmap(ptr,len) UnmapViewOfFile(ptr)
167 #ifdef MDB_USE_POSIX_SEM
169 #define LOCK_MUTEX_R(env) mdb_sem_wait((env)->me_rmutex)
170 #define UNLOCK_MUTEX_R(env) sem_post((env)->me_rmutex)
171 #define LOCK_MUTEX_W(env) mdb_sem_wait((env)->me_wmutex)
172 #define UNLOCK_MUTEX_W(env) sem_post((env)->me_wmutex)
175 mdb_sem_wait(sem_t *sem)
178 while ((rc = sem_wait(sem)) && (rc = errno) == EINTR) ;
183 /** Lock the reader mutex.
185 #define LOCK_MUTEX_R(env) pthread_mutex_lock(&(env)->me_txns->mti_mutex)
186 /** Unlock the reader mutex.
188 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock(&(env)->me_txns->mti_mutex)
190 /** Lock the writer mutex.
191 * Only a single write transaction is allowed at a time. Other writers
192 * will block waiting for this mutex.
194 #define LOCK_MUTEX_W(env) pthread_mutex_lock(&(env)->me_txns->mti_wmutex)
195 /** Unlock the writer mutex.
197 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock(&(env)->me_txns->mti_wmutex)
198 #endif /* MDB_USE_POSIX_SEM */
200 /** Get the error code for the last failed system function.
202 #define ErrCode() errno
204 /** An abstraction for a file handle.
205 * On POSIX systems file handles are small integers. On Windows
206 * they're opaque pointers.
210 /** A value for an invalid file handle.
211 * Mainly used to initialize file variables and signify that they are
214 #define INVALID_HANDLE_VALUE (-1)
216 /** Get the size of a memory page for the system.
217 * This is the basic size that the platform's memory manager uses, and is
218 * fundamental to the use of memory-mapped files.
220 #define GET_PAGESIZE(x) ((x) = sysconf(_SC_PAGE_SIZE))
223 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
226 #define MNAME_LEN (sizeof(pthread_mutex_t))
232 /** A flag for opening a file and requesting synchronous data writes.
233 * This is only used when writing a meta page. It's not strictly needed;
234 * we could just do a normal write and then immediately perform a flush.
235 * But if this flag is available it saves us an extra system call.
237 * @note If O_DSYNC is undefined but exists in /usr/include,
238 * preferably set some compiler flag to get the definition.
239 * Otherwise compile with the less efficient -DMDB_DSYNC=O_SYNC.
242 # define MDB_DSYNC O_DSYNC
246 /** Function for flushing the data of a file. Define this to fsync
247 * if fdatasync() is not supported.
249 #ifndef MDB_FDATASYNC
250 # define MDB_FDATASYNC fdatasync
254 # define MDB_MSYNC(addr,len,flags) msync(addr,len,flags)
265 /** A page number in the database.
266 * Note that 64 bit page numbers are overkill, since pages themselves
267 * already represent 12-13 bits of addressable memory, and the OS will
268 * always limit applications to a maximum of 63 bits of address space.
270 * @note In the #MDB_node structure, we only store 48 bits of this value,
271 * which thus limits us to only 60 bits of addressable data.
273 typedef MDB_ID pgno_t;
275 /** A transaction ID.
276 * See struct MDB_txn.mt_txnid for details.
278 typedef MDB_ID txnid_t;
280 /** @defgroup debug Debug Macros
284 /** Enable debug output.
285 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
286 * read from and written to the database (used for free space management).
291 #if !(__STDC_VERSION__ >= 199901L || defined(__GNUC__))
294 # define DPRINTF (void) /* Vararg macros may be unsupported */
296 static int mdb_debug;
297 static txnid_t mdb_debug_start;
299 /** Print a debug message with printf formatting. */
300 # define DPRINTF(fmt, ...) /**< Requires 2 or more args */ \
301 ((void) ((mdb_debug) && \
302 fprintf(stderr, "%s:%d " fmt "\n", __func__, __LINE__, __VA_ARGS__)))
304 # define DPRINTF(fmt, ...) ((void) 0)
305 # define MDB_DEBUG_SKIP
307 /** Print a debug string.
308 * The string is printed literally, with no format processing.
310 #define DPUTS(arg) DPRINTF("%s", arg)
313 /** A default memory page size.
314 * The actual size is platform-dependent, but we use this for
315 * boot-strapping. We probably should not be using this any more.
316 * The #GET_PAGESIZE() macro is used to get the actual size.
318 * Note that we don't currently support Huge pages. On Linux,
319 * regular data files cannot use Huge pages, and in general
320 * Huge pages aren't actually pageable. We rely on the OS
321 * demand-pager to read our data and page it out when memory
322 * pressure from other processes is high. So until OSs have
323 * actual paging support for Huge pages, they're not viable.
325 #define MDB_PAGESIZE 4096
327 /** The minimum number of keys required in a database page.
328 * Setting this to a larger value will place a smaller bound on the
329 * maximum size of a data item. Data items larger than this size will
330 * be pushed into overflow pages instead of being stored directly in
331 * the B-tree node. This value used to default to 4. With a page size
332 * of 4096 bytes that meant that any item larger than 1024 bytes would
333 * go into an overflow page. That also meant that on average 2-3KB of
334 * each overflow page was wasted space. The value cannot be lower than
335 * 2 because then there would no longer be a tree structure. With this
336 * value, items larger than 2KB will go into overflow pages, and on
337 * average only 1KB will be wasted.
339 #define MDB_MINKEYS 2
341 /** A stamp that identifies a file as an MDB file.
342 * There's nothing special about this value other than that it is easily
343 * recognizable, and it will reflect any byte order mismatches.
345 #define MDB_MAGIC 0xBEEFC0DE
347 /** The version number for a database's file format. */
348 #define MDB_VERSION 1
350 /** @brief The maximum size of a key in the database.
352 * The library rejects bigger keys, and cannot deal with records
353 * with bigger keys stored by a library with bigger max keysize.
355 * We require that keys all fit onto a regular page. This limit
356 * could be raised a bit further if needed; to something just
357 * under #MDB_PAGESIZE / #MDB_MINKEYS.
359 * Note that data items in an #MDB_DUPSORT database are actually keys
360 * of a subDB, so they're also limited to this size.
362 #ifndef MDB_MAXKEYSIZE
363 #define MDB_MAXKEYSIZE 511
366 /** @brief The maximum size of a data item.
368 * We only store a 32 bit value for node sizes.
370 #define MAXDATASIZE 0xffffffffUL
375 * This is used for printing a hex dump of a key's contents.
377 #define DKBUF char kbuf[(MDB_MAXKEYSIZE*2+1)]
378 /** Display a key in hex.
380 * Invoke a function to display a key in hex.
382 #define DKEY(x) mdb_dkey(x, kbuf)
384 #define DKBUF typedef int dummy_kbuf /* so we can put ';' after */
388 /** An invalid page number.
389 * Mainly used to denote an empty tree.
391 #define P_INVALID (~(pgno_t)0)
393 /** Test if the flags \b f are set in a flag word \b w. */
394 #define F_ISSET(w, f) (((w) & (f)) == (f))
396 /** Used for offsets within a single page.
397 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
400 typedef uint16_t indx_t;
402 /** Default size of memory map.
403 * This is certainly too small for any actual applications. Apps should always set
404 * the size explicitly using #mdb_env_set_mapsize().
406 #define DEFAULT_MAPSIZE 1048576
408 /** @defgroup readers Reader Lock Table
409 * Readers don't acquire any locks for their data access. Instead, they
410 * simply record their transaction ID in the reader table. The reader
411 * mutex is needed just to find an empty slot in the reader table. The
412 * slot's address is saved in thread-specific data so that subsequent read
413 * transactions started by the same thread need no further locking to proceed.
415 * If #MDB_NOTLS is set, the slot address is not saved in thread-specific data.
417 * No reader table is used if the database is on a read-only filesystem.
419 * Since the database uses multi-version concurrency control, readers don't
420 * actually need any locking. This table is used to keep track of which
421 * readers are using data from which old transactions, so that we'll know
422 * when a particular old transaction is no longer in use. Old transactions
423 * that have discarded any data pages can then have those pages reclaimed
424 * for use by a later write transaction.
426 * The lock table is constructed such that reader slots are aligned with the
427 * processor's cache line size. Any slot is only ever used by one thread.
428 * This alignment guarantees that there will be no contention or cache
429 * thrashing as threads update their own slot info, and also eliminates
430 * any need for locking when accessing a slot.
432 * A writer thread will scan every slot in the table to determine the oldest
433 * outstanding reader transaction. Any freed pages older than this will be
434 * reclaimed by the writer. The writer doesn't use any locks when scanning
435 * this table. This means that there's no guarantee that the writer will
436 * see the most up-to-date reader info, but that's not required for correct
437 * operation - all we need is to know the upper bound on the oldest reader,
438 * we don't care at all about the newest reader. So the only consequence of
439 * reading stale information here is that old pages might hang around a
440 * while longer before being reclaimed. That's actually good anyway, because
441 * the longer we delay reclaiming old pages, the more likely it is that a
442 * string of contiguous pages can be found after coalescing old pages from
443 * many old transactions together.
446 /** Number of slots in the reader table.
447 * This value was chosen somewhat arbitrarily. 126 readers plus a
448 * couple mutexes fit exactly into 8KB on my development machine.
449 * Applications should set the table size using #mdb_env_set_maxreaders().
451 #define DEFAULT_READERS 126
453 /** The size of a CPU cache line in bytes. We want our lock structures
454 * aligned to this size to avoid false cache line sharing in the
456 * This value works for most CPUs. For Itanium this should be 128.
462 /** The information we store in a single slot of the reader table.
463 * In addition to a transaction ID, we also record the process and
464 * thread ID that owns a slot, so that we can detect stale information,
465 * e.g. threads or processes that went away without cleaning up.
466 * @note We currently don't check for stale records. We simply re-init
467 * the table when we know that we're the only process opening the
470 typedef struct MDB_rxbody {
471 /** Current Transaction ID when this transaction began, or (txnid_t)-1.
472 * Multiple readers that start at the same time will probably have the
473 * same ID here. Again, it's not important to exclude them from
474 * anything; all we need to know is which version of the DB they
475 * started from so we can avoid overwriting any data used in that
476 * particular version.
479 /** The process ID of the process owning this reader txn. */
481 /** The thread ID of the thread owning this txn. */
485 /** The actual reader record, with cacheline padding. */
486 typedef struct MDB_reader {
489 /** shorthand for mrb_txnid */
490 #define mr_txnid mru.mrx.mrb_txnid
491 #define mr_pid mru.mrx.mrb_pid
492 #define mr_tid mru.mrx.mrb_tid
493 /** cache line alignment */
494 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
498 /** The header for the reader table.
499 * The table resides in a memory-mapped file. (This is a different file
500 * than is used for the main database.)
502 * For POSIX the actual mutexes reside in the shared memory of this
503 * mapped file. On Windows, mutexes are named objects allocated by the
504 * kernel; we store the mutex names in this mapped file so that other
505 * processes can grab them. This same approach is also used on
506 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
507 * process-shared POSIX mutexes. For these cases where a named object
508 * is used, the object name is derived from a 64 bit FNV hash of the
509 * environment pathname. As such, naming collisions are extremely
510 * unlikely. If a collision occurs, the results are unpredictable.
512 typedef struct MDB_txbody {
513 /** Stamp identifying this as an MDB file. It must be set
516 /** Version number of this lock file. Must be set to #MDB_VERSION. */
517 uint32_t mtb_version;
518 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
519 char mtb_rmname[MNAME_LEN];
521 /** Mutex protecting access to this table.
522 * This is the reader lock that #LOCK_MUTEX_R acquires.
524 pthread_mutex_t mtb_mutex;
526 /** The ID of the last transaction committed to the database.
527 * This is recorded here only for convenience; the value can always
528 * be determined by reading the main database meta pages.
531 /** The number of slots that have been used in the reader table.
532 * This always records the maximum count, it is not decremented
533 * when readers release their slots.
535 unsigned mtb_numreaders;
538 /** The actual reader table definition. */
539 typedef struct MDB_txninfo {
542 #define mti_magic mt1.mtb.mtb_magic
543 #define mti_version mt1.mtb.mtb_version
544 #define mti_mutex mt1.mtb.mtb_mutex
545 #define mti_rmname mt1.mtb.mtb_rmname
546 #define mti_txnid mt1.mtb.mtb_txnid
547 #define mti_numreaders mt1.mtb.mtb_numreaders
548 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
551 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
552 char mt2_wmname[MNAME_LEN];
553 #define mti_wmname mt2.mt2_wmname
555 pthread_mutex_t mt2_wmutex;
556 #define mti_wmutex mt2.mt2_wmutex
558 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
560 MDB_reader mti_readers[1];
564 /** Common header for all page types.
565 * Overflow records occupy a number of contiguous pages with no
566 * headers on any page after the first.
568 typedef struct MDB_page {
569 #define mp_pgno mp_p.p_pgno
570 #define mp_next mp_p.p_next
572 pgno_t p_pgno; /**< page number */
573 void * p_next; /**< for in-memory list of freed structs */
576 /** @defgroup mdb_page Page Flags
578 * Flags for the page headers.
581 #define P_BRANCH 0x01 /**< branch page */
582 #define P_LEAF 0x02 /**< leaf page */
583 #define P_OVERFLOW 0x04 /**< overflow page */
584 #define P_META 0x08 /**< meta page */
585 #define P_DIRTY 0x10 /**< dirty page */
586 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
587 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
588 #define P_KEEP 0x8000 /**< leave this page alone during spill */
590 uint16_t mp_flags; /**< @ref mdb_page */
591 #define mp_lower mp_pb.pb.pb_lower
592 #define mp_upper mp_pb.pb.pb_upper
593 #define mp_pages mp_pb.pb_pages
596 indx_t pb_lower; /**< lower bound of free space */
597 indx_t pb_upper; /**< upper bound of free space */
599 uint32_t pb_pages; /**< number of overflow pages */
601 indx_t mp_ptrs[1]; /**< dynamic size */
604 /** Size of the page header, excluding dynamic data at the end */
605 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
607 /** Address of first usable data byte in a page, after the header */
608 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
610 /** Number of nodes on a page */
611 #define NUMKEYS(p) (((p)->mp_lower - PAGEHDRSZ) >> 1)
613 /** The amount of space remaining in the page */
614 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
616 /** The percentage of space used in the page, in tenths of a percent. */
617 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
618 ((env)->me_psize - PAGEHDRSZ))
619 /** The minimum page fill factor, in tenths of a percent.
620 * Pages emptier than this are candidates for merging.
622 #define FILL_THRESHOLD 250
624 /** Test if a page is a leaf page */
625 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
626 /** Test if a page is a LEAF2 page */
627 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
628 /** Test if a page is a branch page */
629 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
630 /** Test if a page is an overflow page */
631 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
632 /** Test if a page is a sub page */
633 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
635 /** The number of overflow pages needed to store the given size. */
636 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
638 /** Header for a single key/data pair within a page.
639 * We guarantee 2-byte alignment for nodes.
641 typedef struct MDB_node {
642 /** lo and hi are used for data size on leaf nodes and for
643 * child pgno on branch nodes. On 64 bit platforms, flags
644 * is also used for pgno. (Branch nodes have no flags).
645 * They are in host byte order in case that lets some
646 * accesses be optimized into a 32-bit word access.
648 #define mn_lo mn_offset[BYTE_ORDER!=LITTLE_ENDIAN]
649 #define mn_hi mn_offset[BYTE_ORDER==LITTLE_ENDIAN] /**< part of dsize or pgno */
650 unsigned short mn_offset[2]; /**< storage for #mn_lo and #mn_hi */
651 /** @defgroup mdb_node Node Flags
653 * Flags for node headers.
656 #define F_BIGDATA 0x01 /**< data put on overflow page */
657 #define F_SUBDATA 0x02 /**< data is a sub-database */
658 #define F_DUPDATA 0x04 /**< data has duplicates */
660 /** valid flags for #mdb_node_add() */
661 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
664 unsigned short mn_flags; /**< @ref mdb_node */
665 unsigned short mn_ksize; /**< key size */
666 char mn_data[1]; /**< key and data are appended here */
669 /** Size of the node header, excluding dynamic data at the end */
670 #define NODESIZE offsetof(MDB_node, mn_data)
672 /** Bit position of top word in page number, for shifting mn_flags */
673 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
675 /** Size of a node in a branch page with a given key.
676 * This is just the node header plus the key, there is no data.
678 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
680 /** Size of a node in a leaf page with a given key and data.
681 * This is node header plus key plus data size.
683 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
685 /** Address of node \b i in page \b p */
686 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i]))
688 /** Address of the key for the node */
689 #define NODEKEY(node) (void *)((node)->mn_data)
691 /** Address of the data for a node */
692 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
694 /** Get the page number pointed to by a branch node */
695 #define NODEPGNO(node) \
696 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
697 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
698 /** Set the page number in a branch node */
699 #define SETPGNO(node,pgno) do { \
700 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
701 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
703 /** Get the size of the data in a leaf node */
704 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
705 /** Set the size of the data for a leaf node */
706 #define SETDSZ(node,size) do { \
707 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
708 /** The size of a key in a node */
709 #define NODEKSZ(node) ((node)->mn_ksize)
711 /** Copy a page number from src to dst */
713 #define COPY_PGNO(dst,src) dst = src
715 #if SIZE_MAX > 4294967295UL
716 #define COPY_PGNO(dst,src) do { \
717 unsigned short *s, *d; \
718 s = (unsigned short *)&(src); \
719 d = (unsigned short *)&(dst); \
726 #define COPY_PGNO(dst,src) do { \
727 unsigned short *s, *d; \
728 s = (unsigned short *)&(src); \
729 d = (unsigned short *)&(dst); \
735 /** The address of a key in a LEAF2 page.
736 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
737 * There are no node headers, keys are stored contiguously.
739 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
741 /** Set the \b node's key into \b key, if requested. */
742 #define MDB_GET_KEY(node, key) { if ((key) != NULL) { \
743 (key)->mv_size = NODEKSZ(node); (key)->mv_data = NODEKEY(node); } }
745 /** Information about a single database in the environment. */
746 typedef struct MDB_db {
747 uint32_t md_pad; /**< also ksize for LEAF2 pages */
748 uint16_t md_flags; /**< @ref mdb_dbi_open */
749 uint16_t md_depth; /**< depth of this tree */
750 pgno_t md_branch_pages; /**< number of internal pages */
751 pgno_t md_leaf_pages; /**< number of leaf pages */
752 pgno_t md_overflow_pages; /**< number of overflow pages */
753 size_t md_entries; /**< number of data items */
754 pgno_t md_root; /**< the root page of this tree */
757 /** mdb_dbi_open flags */
758 #define MDB_VALID 0x8000 /**< DB handle is valid, for me_dbflags */
759 #define PERSISTENT_FLAGS (0xffff & ~(MDB_VALID))
760 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
761 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
763 /** Handle for the DB used to track free pages. */
765 /** Handle for the default DB. */
768 /** Meta page content. */
769 typedef struct MDB_meta {
770 /** Stamp identifying this as an MDB file. It must be set
773 /** Version number of this lock file. Must be set to #MDB_VERSION. */
775 void *mm_address; /**< address for fixed mapping */
776 size_t mm_mapsize; /**< size of mmap region */
777 MDB_db mm_dbs[2]; /**< first is free space, 2nd is main db */
778 /** The size of pages used in this DB */
779 #define mm_psize mm_dbs[0].md_pad
780 /** Any persistent environment flags. @ref mdb_env */
781 #define mm_flags mm_dbs[0].md_flags
782 pgno_t mm_last_pg; /**< last used page in file */
783 txnid_t mm_txnid; /**< txnid that committed this page */
786 /** Buffer for a stack-allocated dirty page.
787 * The members define size and alignment, and silence type
788 * aliasing warnings. They are not used directly; that could
789 * mean incorrectly using several union members in parallel.
791 typedef union MDB_pagebuf {
792 char mb_raw[MDB_PAGESIZE];
795 char mm_pad[PAGEHDRSZ];
800 /** Auxiliary DB info.
801 * The information here is mostly static/read-only. There is
802 * only a single copy of this record in the environment.
804 typedef struct MDB_dbx {
805 MDB_val md_name; /**< name of the database */
806 MDB_cmp_func *md_cmp; /**< function for comparing keys */
807 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
808 MDB_rel_func *md_rel; /**< user relocate function */
809 void *md_relctx; /**< user-provided context for md_rel */
812 /** A database transaction.
813 * Every operation requires a transaction handle.
816 MDB_txn *mt_parent; /**< parent of a nested txn */
817 MDB_txn *mt_child; /**< nested txn under this txn */
818 pgno_t mt_next_pgno; /**< next unallocated page */
819 /** The ID of this transaction. IDs are integers incrementing from 1.
820 * Only committed write transactions increment the ID. If a transaction
821 * aborts, the ID may be re-used by the next writer.
824 MDB_env *mt_env; /**< the DB environment */
825 /** The list of pages that became unused during this transaction.
828 /** The list of dirty pages we temporarily wrote to disk
829 * because the dirty list was full.
831 MDB_IDL mt_spill_pgs;
833 MDB_ID2L dirty_list; /**< for write txns: modified pages */
834 MDB_reader *reader; /**< this thread's reader table slot or NULL */
836 /** Array of records for each DB known in the environment. */
838 /** Array of MDB_db records for each known DB */
840 /** @defgroup mt_dbflag Transaction DB Flags
844 #define DB_DIRTY 0x01 /**< DB was written in this txn */
845 #define DB_STALE 0x02 /**< DB record is older than txnID */
846 #define DB_NEW 0x04 /**< DB handle opened in this txn */
847 #define DB_VALID 0x08 /**< DB handle is valid, see also #MDB_VALID */
849 /** In write txns, array of cursors for each DB */
850 MDB_cursor **mt_cursors;
851 /** Array of flags for each DB */
852 unsigned char *mt_dbflags;
853 /** Number of DB records in use. This number only ever increments;
854 * we don't decrement it when individual DB handles are closed.
858 /** @defgroup mdb_txn Transaction Flags
862 #define MDB_TXN_RDONLY 0x01 /**< read-only transaction */
863 #define MDB_TXN_ERROR 0x02 /**< an error has occurred */
864 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
865 #define MDB_TXN_SPILLS 0x08 /**< txn or a parent has spilled pages */
867 unsigned int mt_flags; /**< @ref mdb_txn */
868 /** dirty_list maxsize - # of allocated pages allowed, including in parent txns */
869 unsigned int mt_dirty_room;
870 /** Tracks which of the two meta pages was used at the start
871 * of this transaction.
873 unsigned int mt_toggle;
876 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
877 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
878 * raise this on a 64 bit machine.
880 #define CURSOR_STACK 32
884 /** Cursors are used for all DB operations */
886 /** Next cursor on this DB in this txn */
888 /** Backup of the original cursor if this cursor is a shadow */
889 MDB_cursor *mc_backup;
890 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
891 struct MDB_xcursor *mc_xcursor;
892 /** The transaction that owns this cursor */
894 /** The database handle this cursor operates on */
896 /** The database record for this cursor */
898 /** The database auxiliary record for this cursor */
900 /** The @ref mt_dbflag for this database */
901 unsigned char *mc_dbflag;
902 unsigned short mc_snum; /**< number of pushed pages */
903 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
904 /** @defgroup mdb_cursor Cursor Flags
906 * Cursor state flags.
909 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
910 #define C_EOF 0x02 /**< No more data */
911 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
912 #define C_SPLITTING 0x20 /**< Cursor is in page_split */
913 #define C_UNTRACK 0x40 /**< Un-track cursor when closing */
915 unsigned int mc_flags; /**< @ref mdb_cursor */
916 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
917 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
920 /** Context for sorted-dup records.
921 * We could have gone to a fully recursive design, with arbitrarily
922 * deep nesting of sub-databases. But for now we only handle these
923 * levels - main DB, optional sub-DB, sorted-duplicate DB.
925 typedef struct MDB_xcursor {
926 /** A sub-cursor for traversing the Dup DB */
927 MDB_cursor mx_cursor;
928 /** The database record for this Dup DB */
930 /** The auxiliary DB record for this Dup DB */
932 /** The @ref mt_dbflag for this Dup DB */
933 unsigned char mx_dbflag;
936 /** State of FreeDB old pages, stored in the MDB_env */
937 typedef struct MDB_pgstate {
938 pgno_t *mf_pghead; /**< Reclaimed freeDB pages, or NULL before use */
939 txnid_t mf_pglast; /**< ID of last used record, or 0 if !mf_pghead */
942 /** The database environment. */
944 HANDLE me_fd; /**< The main data file */
945 HANDLE me_lfd; /**< The lock file */
946 HANDLE me_mfd; /**< just for writing the meta pages */
947 /** Failed to update the meta page. Probably an I/O error. */
948 #define MDB_FATAL_ERROR 0x80000000U
949 /** Some fields are initialized. */
950 #define MDB_ENV_ACTIVE 0x20000000U
951 /** me_txkey is set */
952 #define MDB_ENV_TXKEY 0x10000000U
953 uint32_t me_flags; /**< @ref mdb_env */
954 unsigned int me_psize; /**< size of a page, from #GET_PAGESIZE */
955 unsigned int me_maxreaders; /**< size of the reader table */
956 unsigned int me_numreaders; /**< max numreaders set by this env */
957 MDB_dbi me_numdbs; /**< number of DBs opened */
958 MDB_dbi me_maxdbs; /**< size of the DB table */
959 pid_t me_pid; /**< process ID of this env */
960 char *me_path; /**< path to the DB files */
961 char *me_map; /**< the memory map of the data file */
962 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
963 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
964 MDB_txn *me_txn; /**< current write transaction */
965 size_t me_mapsize; /**< size of the data memory map */
966 off_t me_size; /**< current file size */
967 pgno_t me_maxpg; /**< me_mapsize / me_psize */
968 MDB_dbx *me_dbxs; /**< array of static DB info */
969 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
970 pthread_key_t me_txkey; /**< thread-key for readers */
971 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
972 # define me_pglast me_pgstate.mf_pglast
973 # define me_pghead me_pgstate.mf_pghead
974 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
975 /** IDL of pages that became unused in a write txn */
977 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */
978 MDB_ID2L me_dirty_list;
979 /** Max number of freelist items that can fit in a single overflow page */
981 /** Max size of a node on a page */
982 unsigned int me_nodemax;
984 HANDLE me_rmutex; /* Windows mutexes don't reside in shared mem */
986 #elif defined(MDB_USE_POSIX_SEM)
987 sem_t *me_rmutex; /* Shared mutexes are not supported */
992 /** Nested transaction */
993 typedef struct MDB_ntxn {
994 MDB_txn mnt_txn; /* the transaction */
995 MDB_pgstate mnt_pgstate; /* parent transaction's saved freestate */
998 /** max number of pages to commit in one writev() call */
999 #define MDB_COMMIT_PAGES 64
1000 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
1001 #undef MDB_COMMIT_PAGES
1002 #define MDB_COMMIT_PAGES IOV_MAX
1005 /* max bytes to write in one call */
1006 #define MAX_WRITE (0x80000000U >> (sizeof(ssize_t) == 4))
1008 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
1009 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
1010 static int mdb_page_touch(MDB_cursor *mc);
1012 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp, int *lvl);
1013 static int mdb_page_search_root(MDB_cursor *mc,
1014 MDB_val *key, int modify);
1015 #define MDB_PS_MODIFY 1
1016 #define MDB_PS_ROOTONLY 2
1017 static int mdb_page_search(MDB_cursor *mc,
1018 MDB_val *key, int flags);
1019 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1021 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1022 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1023 pgno_t newpgno, unsigned int nflags);
1025 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1026 static int mdb_env_pick_meta(const MDB_env *env);
1027 static int mdb_env_write_meta(MDB_txn *txn);
1028 #if !(defined(_WIN32) || defined(MDB_USE_POSIX_SEM)) /* Drop unused excl arg */
1029 # define mdb_env_close0(env, excl) mdb_env_close1(env)
1031 static void mdb_env_close0(MDB_env *env, int excl);
1033 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1034 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1035 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1036 static void mdb_node_del(MDB_page *mp, indx_t indx, int ksize);
1037 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1038 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
1039 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
1040 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1041 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1043 static int mdb_rebalance(MDB_cursor *mc);
1044 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1046 static void mdb_cursor_pop(MDB_cursor *mc);
1047 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1049 static int mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf);
1050 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1051 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1052 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1053 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1055 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1056 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1058 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1059 static void mdb_xcursor_init0(MDB_cursor *mc);
1060 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1062 static int mdb_drop0(MDB_cursor *mc, int subs);
1063 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1066 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1070 static SECURITY_DESCRIPTOR mdb_null_sd;
1071 static SECURITY_ATTRIBUTES mdb_all_sa;
1072 static int mdb_sec_inited;
1075 /** Return the library version info. */
1077 mdb_version(int *major, int *minor, int *patch)
1079 if (major) *major = MDB_VERSION_MAJOR;
1080 if (minor) *minor = MDB_VERSION_MINOR;
1081 if (patch) *patch = MDB_VERSION_PATCH;
1082 return MDB_VERSION_STRING;
1085 /** Table of descriptions for MDB @ref errors */
1086 static char *const mdb_errstr[] = {
1087 "MDB_KEYEXIST: Key/data pair already exists",
1088 "MDB_NOTFOUND: No matching key/data pair found",
1089 "MDB_PAGE_NOTFOUND: Requested page not found",
1090 "MDB_CORRUPTED: Located page was wrong type",
1091 "MDB_PANIC: Update of meta page failed",
1092 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1093 "MDB_INVALID: File is not an MDB file",
1094 "MDB_MAP_FULL: Environment mapsize limit reached",
1095 "MDB_DBS_FULL: Environment maxdbs limit reached",
1096 "MDB_READERS_FULL: Environment maxreaders limit reached",
1097 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1098 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1099 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1100 "MDB_PAGE_FULL: Internal error - page has no more space",
1101 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1102 "MDB_INCOMPATIBLE: Database flags changed or would change",
1103 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1107 mdb_strerror(int err)
1111 return ("Successful return: 0");
1113 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1114 i = err - MDB_KEYEXIST;
1115 return mdb_errstr[i];
1118 return strerror(err);
1122 /** Display a key in hexadecimal and return the address of the result.
1123 * @param[in] key the key to display
1124 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1125 * @return The key in hexadecimal form.
1128 mdb_dkey(MDB_val *key, char *buf)
1131 unsigned char *c = key->mv_data;
1137 if (key->mv_size > MDB_MAXKEYSIZE)
1138 return "MDB_MAXKEYSIZE";
1139 /* may want to make this a dynamic check: if the key is mostly
1140 * printable characters, print it as-is instead of converting to hex.
1144 for (i=0; i<key->mv_size; i++)
1145 ptr += sprintf(ptr, "%02x", *c++);
1147 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1152 /** Display all the keys in the page. */
1154 mdb_page_list(MDB_page *mp)
1157 unsigned int i, nkeys, nsize;
1161 nkeys = NUMKEYS(mp);
1162 fprintf(stderr, "Page %zu numkeys %d\n", mp->mp_pgno, nkeys);
1163 for (i=0; i<nkeys; i++) {
1164 node = NODEPTR(mp, i);
1165 key.mv_size = node->mn_ksize;
1166 key.mv_data = node->mn_data;
1167 nsize = NODESIZE + NODEKSZ(node) + sizeof(indx_t);
1168 if (IS_BRANCH(mp)) {
1169 fprintf(stderr, "key %d: page %zu, %s\n", i, NODEPGNO(node),
1172 if (F_ISSET(node->mn_flags, F_BIGDATA))
1173 nsize += sizeof(pgno_t);
1175 nsize += NODEDSZ(node);
1176 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1182 mdb_cursor_chk(MDB_cursor *mc)
1188 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1189 for (i=0; i<mc->mc_top; i++) {
1191 node = NODEPTR(mp, mc->mc_ki[i]);
1192 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1195 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1201 /** Count all the pages in each DB and in the freelist
1202 * and make sure it matches the actual number of pages
1205 static void mdb_audit(MDB_txn *txn)
1209 MDB_ID freecount, count;
1214 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1215 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1216 freecount += *(MDB_ID *)data.mv_data;
1219 for (i = 0; i<txn->mt_numdbs; i++) {
1221 mdb_cursor_init(&mc, txn, i, &mx);
1222 if (txn->mt_dbs[i].md_root == P_INVALID)
1224 count += txn->mt_dbs[i].md_branch_pages +
1225 txn->mt_dbs[i].md_leaf_pages +
1226 txn->mt_dbs[i].md_overflow_pages;
1227 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1228 mdb_page_search(&mc, NULL, 0);
1232 mp = mc.mc_pg[mc.mc_top];
1233 for (j=0; j<NUMKEYS(mp); j++) {
1234 MDB_node *leaf = NODEPTR(mp, j);
1235 if (leaf->mn_flags & F_SUBDATA) {
1237 memcpy(&db, NODEDATA(leaf), sizeof(db));
1238 count += db.md_branch_pages + db.md_leaf_pages +
1239 db.md_overflow_pages;
1243 while (mdb_cursor_sibling(&mc, 1) == 0);
1246 if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
1247 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1248 txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
1254 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1256 return txn->mt_dbxs[dbi].md_cmp(a, b);
1260 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1262 return txn->mt_dbxs[dbi].md_dcmp(a, b);
1265 /** Allocate a page.
1266 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1269 mdb_page_malloc(MDB_txn *txn, unsigned num)
1271 MDB_env *env = txn->mt_env;
1272 MDB_page *ret = env->me_dpages;
1273 size_t sz = env->me_psize;
1276 VGMEMP_ALLOC(env, ret, sz);
1277 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1278 env->me_dpages = ret->mp_next;
1284 if ((ret = malloc(sz)) != NULL) {
1285 VGMEMP_ALLOC(env, ret, sz);
1290 /** Free a single page.
1291 * Saves single pages to a list, for future reuse.
1292 * (This is not used for multi-page overflow pages.)
1295 mdb_page_free(MDB_env *env, MDB_page *mp)
1297 mp->mp_next = env->me_dpages;
1298 VGMEMP_FREE(env, mp);
1299 env->me_dpages = mp;
1302 /* Free a dirty page */
1304 mdb_dpage_free(MDB_env *env, MDB_page *dp)
1306 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1307 mdb_page_free(env, dp);
1309 /* large pages just get freed directly */
1310 VGMEMP_FREE(env, dp);
1315 /** Return all dirty pages to dpage list */
1317 mdb_dlist_free(MDB_txn *txn)
1319 MDB_env *env = txn->mt_env;
1320 MDB_ID2L dl = txn->mt_u.dirty_list;
1321 unsigned i, n = dl[0].mid;
1323 for (i = 1; i <= n; i++) {
1324 mdb_dpage_free(env, dl[i].mptr);
1329 /* Set or clear P_KEEP in non-overflow, non-sub pages in known cursors.
1330 * When clearing, only consider backup cursors (from parent txns) since
1331 * other P_KEEP flags have already been cleared.
1332 * @param[in] mc A cursor handle for the current operation.
1333 * @param[in] pflags Flags of the pages to update:
1334 * P_DIRTY to set P_KEEP, P_DIRTY|P_KEEP to clear it.
1337 mdb_cursorpages_mark(MDB_cursor *mc, unsigned pflags)
1339 MDB_txn *txn = mc->mc_txn;
1340 MDB_cursor *m2, *m3;
1344 if (mc->mc_flags & C_UNTRACK)
1345 mc = NULL; /* will find mc in mt_cursors */
1346 for (i = txn->mt_numdbs;; mc = txn->mt_cursors[--i]) {
1347 for (; mc; mc=mc->mc_next) {
1348 m2 = pflags == P_DIRTY ? mc : mc->mc_backup;
1349 for (; m2; m2 = m2->mc_backup) {
1350 for (m3=m2; m3->mc_flags & C_INITIALIZED; m3=&mx->mx_cursor) {
1351 for (j=0; j<m3->mc_snum; j++)
1352 if ((m3->mc_pg[j]->mp_flags & (P_SUBP|P_DIRTY|P_KEEP))
1354 m3->mc_pg[j]->mp_flags ^= P_KEEP;
1355 if (!(m3->mc_db->md_flags & MDB_DUPSORT))
1357 /* Cursor backups have mx malloced at the end of m2 */
1358 mx = (m3 == mc ? m3->mc_xcursor : (MDB_xcursor *)(m3+1));
1367 static int mdb_page_flush(MDB_txn *txn);
1369 /** Spill pages from the dirty list back to disk.
1370 * This is intended to prevent running into #MDB_TXN_FULL situations,
1371 * but note that they may still occur in a few cases:
1372 * 1) pages in #MDB_DUPSORT sub-DBs are never spilled, so if there
1373 * are too many of these dirtied in one txn, the txn may still get
1375 * 2) child txns may run out of space if their parents dirtied a
1376 * lot of pages and never spilled them. TODO: we probably should do
1377 * a preemptive spill during #mdb_txn_begin() of a child txn, if
1378 * the parent's dirty_room is below a given threshold.
1379 * 3) our estimate of the txn size could be too small. At the
1380 * moment this seems unlikely.
1382 * Otherwise, if not using nested txns, it is expected that apps will
1383 * not run into #MDB_TXN_FULL any more. The pages are flushed to disk
1384 * the same way as for a txn commit, e.g. their P_DIRTY flag is cleared.
1385 * If the txn never references them again, they can be left alone.
1386 * If the txn only reads them, they can be used without any fuss.
1387 * If the txn writes them again, they can be dirtied immediately without
1388 * going thru all of the work of #mdb_page_touch(). Such references are
1389 * handled by #mdb_page_unspill().
1391 * Also note, we never spill DB root pages, nor pages of active cursors,
1392 * because we'll need these back again soon anyway. And in nested txns,
1393 * we can't spill a page in a child txn if it was already spilled in a
1394 * parent txn. That would alter the parent txns' data even though
1395 * the child hasn't committed yet, and we'd have no way to undo it if
1396 * the child aborted.
1398 * @param[in] m0 cursor A cursor handle identifying the transaction and
1399 * database for which we are checking space.
1400 * @param[in] key For a put operation, the key being stored.
1401 * @param[in] data For a put operation, the data being stored.
1402 * @return 0 on success, non-zero on failure.
1405 mdb_page_spill(MDB_cursor *m0, MDB_val *key, MDB_val *data)
1407 MDB_txn *txn = m0->mc_txn;
1409 MDB_ID2L dl = txn->mt_u.dirty_list;
1413 if (m0->mc_flags & C_SUB)
1416 /* Estimate how much space this op will take */
1417 i = m0->mc_db->md_depth;
1418 /* Named DBs also dirty the main DB */
1419 if (m0->mc_dbi > MAIN_DBI)
1420 i += txn->mt_dbs[MAIN_DBI].md_depth;
1421 /* For puts, roughly factor in the key+data size */
1423 i += (LEAFSIZE(key, data) + txn->mt_env->me_psize) / txn->mt_env->me_psize;
1424 i += i; /* double it for good measure */
1426 if (txn->mt_dirty_room > i)
1429 if (!txn->mt_spill_pgs) {
1430 txn->mt_spill_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX);
1431 if (!txn->mt_spill_pgs)
1435 /* Mark all the dirty root pages we want to preserve */
1436 for (i=0; i<txn->mt_numdbs; i++) {
1437 if (txn->mt_dbflags[i] & DB_DIRTY) {
1438 j = mdb_mid2l_search(dl, txn->mt_dbs[i].md_root);
1439 if (j <= dl[0].mid) {
1441 dp->mp_flags |= P_KEEP;
1446 /* Preserve pages used by cursors */
1447 mdb_cursorpages_mark(m0, P_DIRTY);
1449 /* Save the page IDs of all the pages we're flushing */
1450 for (i=1; i<=dl[0].mid; i++) {
1452 if (dp->mp_flags & P_KEEP)
1454 /* Can't spill twice, make sure it's not already in a parent's
1457 if (txn->mt_parent) {
1459 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
1460 if (tx2->mt_spill_pgs) {
1461 j = mdb_midl_search(tx2->mt_spill_pgs, dl[i].mid);
1462 if (j <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[j] == dl[i].mid) {
1463 dp->mp_flags |= P_KEEP;
1471 if ((rc = mdb_midl_append(&txn->mt_spill_pgs, dl[i].mid)))
1474 mdb_midl_sort(txn->mt_spill_pgs);
1476 rc = mdb_page_flush(txn);
1478 mdb_cursorpages_mark(m0, P_DIRTY|P_KEEP);
1481 txn->mt_dirty_room = MDB_IDL_UM_MAX - dl[0].mid;
1482 txn->mt_flags |= MDB_TXN_SPILLS;
1487 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
1489 mdb_find_oldest(MDB_txn *txn)
1492 txnid_t mr, oldest = txn->mt_txnid - 1;
1493 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
1494 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
1504 /** Add a page to the txn's dirty list */
1506 mdb_page_dirty(MDB_txn *txn, MDB_page *mp)
1509 int (*insert)(MDB_ID2L, MDB_ID2 *);
1511 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1512 insert = mdb_mid2l_append;
1514 insert = mdb_mid2l_insert;
1516 mid.mid = mp->mp_pgno;
1518 insert(txn->mt_u.dirty_list, &mid);
1519 txn->mt_dirty_room--;
1522 /** Allocate pages for writing.
1523 * If there are free pages available from older transactions, they
1524 * will be re-used first. Otherwise a new page will be allocated.
1525 * @param[in] mc cursor A cursor handle identifying the transaction and
1526 * database for which we are allocating.
1527 * @param[in] num the number of pages to allocate.
1528 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
1529 * will always be satisfied by a single contiguous chunk of memory.
1530 * @return 0 on success, non-zero on failure.
1533 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
1535 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
1536 /* Get at most <Max_retries> more freeDB records once me_pghead
1537 * has enough pages. If not enough, use new pages from the map.
1538 * If <Paranoid> and mc is updating the freeDB, only get new
1539 * records if me_pghead is empty. Then the freelist cannot play
1540 * catch-up with itself by growing while trying to save it.
1542 enum { Paranoid = 1, Max_retries = 500 };
1544 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ };
1546 int rc, n2 = num-1, retry = Max_retries;
1547 MDB_txn *txn = mc->mc_txn;
1548 MDB_env *env = txn->mt_env;
1549 pgno_t pgno, *mop = env->me_pghead;
1550 unsigned i, j, k, mop_len = mop ? mop[0] : 0;
1552 txnid_t oldest = 0, last;
1558 /* If our dirty list is already full, we can't do anything */
1559 if (txn->mt_dirty_room == 0)
1560 return MDB_TXN_FULL;
1562 for (op = MDB_FIRST;; op = MDB_NEXT) {
1565 pgno_t *idl, old_id, new_id;
1567 /* Seek a big enough contiguous page range. Prefer
1568 * pages at the tail, just truncating the list.
1570 if (mop_len >= (unsigned)num) {
1574 if (mop[i-n2] == pgno+n2)
1576 } while (--i >= (unsigned)num);
1577 if (Max_retries < INT_MAX && --retry < 0)
1581 if (op == MDB_FIRST) { /* 1st iteration */
1582 /* Prepare to fetch more and coalesce */
1583 oldest = mdb_find_oldest(txn);
1584 last = env->me_pglast;
1585 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1588 key.mv_data = &last; /* will loop up last+1 */
1589 key.mv_size = sizeof(last);
1591 if (Paranoid && mc->mc_dbi == FREE_DBI)
1594 if (Paranoid && retry < 0 && mop_len)
1598 /* Do not fetch more if the record will be too recent */
1601 rc = mdb_cursor_get(&m2, &key, NULL, op);
1603 if (rc == MDB_NOTFOUND)
1607 last = *(txnid_t*)key.mv_data;
1610 np = m2.mc_pg[m2.mc_top];
1611 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
1612 if ((rc = mdb_node_read(txn, leaf, &data)) != MDB_SUCCESS)
1615 idl = (MDB_ID *) data.mv_data;
1618 if (!(env->me_pghead = mop = mdb_midl_alloc(i)))
1621 if ((rc = mdb_midl_need(&env->me_pghead, i)) != 0)
1623 mop = env->me_pghead;
1625 env->me_pglast = last;
1627 DPRINTF("IDL read txn %zu root %zu num %u",
1628 last, txn->mt_dbs[FREE_DBI].md_root, i);
1630 DPRINTF("IDL %zu", idl[k]);
1632 /* Merge in descending sorted order */
1635 mop[0] = (pgno_t)-1;
1639 for (; old_id < new_id; old_id = mop[--j])
1646 /* Use new pages from the map when nothing suitable in the freeDB */
1648 pgno = txn->mt_next_pgno;
1649 if (pgno + num >= env->me_maxpg) {
1650 DPUTS("DB size maxed out");
1651 return MDB_MAP_FULL;
1655 if (env->me_flags & MDB_WRITEMAP) {
1656 np = (MDB_page *)(env->me_map + env->me_psize * pgno);
1658 if (!(np = mdb_page_malloc(txn, num)))
1662 mop[0] = mop_len -= num;
1663 /* Move any stragglers down */
1664 for (j = i-num; j < mop_len; )
1665 mop[++j] = mop[++i];
1667 txn->mt_next_pgno = pgno + num;
1670 mdb_page_dirty(txn, np);
1676 /** Copy the used portions of a non-overflow page.
1677 * @param[in] dst page to copy into
1678 * @param[in] src page to copy from
1679 * @param[in] psize size of a page
1682 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
1684 enum { Align = sizeof(pgno_t) };
1685 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
1687 /* If page isn't full, just copy the used portion. Adjust
1688 * alignment so memcpy may copy words instead of bytes.
1690 if ((unused &= -Align) && !IS_LEAF2(src)) {
1692 memcpy(dst, src, (lower + (Align-1)) & -Align);
1693 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper),
1696 memcpy(dst, src, psize - unused);
1700 /** Pull a page off the txn's spill list, if present.
1701 * If a page being referenced was spilled to disk in this txn, bring
1702 * it back and make it dirty/writable again.
1703 * @param[in] tx0 the transaction handle.
1704 * @param[in] mp the page being referenced.
1705 * @param[out] ret the writable page, if any. ret is unchanged if
1706 * mp wasn't spilled.
1709 mdb_page_unspill(MDB_txn *tx0, MDB_page *mp, MDB_page **ret)
1711 MDB_env *env = tx0->mt_env;
1714 pgno_t pgno = mp->mp_pgno;
1716 for (txn = tx0; txn; txn=txn->mt_parent) {
1717 if (!txn->mt_spill_pgs)
1719 x = mdb_midl_search(txn->mt_spill_pgs, pgno);
1720 if (x <= txn->mt_spill_pgs[0] && txn->mt_spill_pgs[x] == pgno) {
1723 if (IS_OVERFLOW(mp))
1727 if (env->me_flags & MDB_WRITEMAP) {
1730 np = mdb_page_malloc(txn, num);
1734 memcpy(np, mp, num * env->me_psize);
1736 mdb_page_copy(np, mp, env->me_psize);
1739 /* If in current txn, this page is no longer spilled */
1740 for (; x < txn->mt_spill_pgs[0]; x++)
1741 txn->mt_spill_pgs[x] = txn->mt_spill_pgs[x+1];
1742 txn->mt_spill_pgs[0]--;
1743 } /* otherwise, if belonging to a parent txn, the
1744 * page remains spilled until child commits
1746 mdb_page_dirty(tx0, np);
1747 np->mp_flags |= P_DIRTY;
1755 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
1756 * @param[in] mc cursor pointing to the page to be touched
1757 * @return 0 on success, non-zero on failure.
1760 mdb_page_touch(MDB_cursor *mc)
1762 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
1763 MDB_txn *txn = mc->mc_txn;
1764 MDB_cursor *m2, *m3;
1769 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
1770 if (txn->mt_flags & MDB_TXN_SPILLS) {
1772 rc = mdb_page_unspill(txn, mp, &np);
1778 if ((rc = mdb_midl_need(&txn->mt_free_pgs, 1)) ||
1779 (rc = mdb_page_alloc(mc, 1, &np)))
1782 DPRINTF("touched db %u page %zu -> %zu", mc->mc_dbi,mp->mp_pgno,pgno);
1783 assert(mp->mp_pgno != pgno);
1784 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
1785 /* Update the parent page, if any, to point to the new page */
1787 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
1788 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
1789 SETPGNO(node, pgno);
1791 mc->mc_db->md_root = pgno;
1793 } else if (txn->mt_parent && !IS_SUBP(mp)) {
1794 MDB_ID2 mid, *dl = txn->mt_u.dirty_list;
1796 /* If txn has a parent, make sure the page is in our
1800 unsigned x = mdb_mid2l_search(dl, pgno);
1801 if (x <= dl[0].mid && dl[x].mid == pgno) {
1802 if (mp != dl[x].mptr) { /* bad cursor? */
1803 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
1804 return MDB_CORRUPTED;
1809 assert(dl[0].mid < MDB_IDL_UM_MAX);
1811 np = mdb_page_malloc(txn, 1);
1816 mdb_mid2l_insert(dl, &mid);
1821 mdb_page_copy(np, mp, txn->mt_env->me_psize);
1823 np->mp_flags |= P_DIRTY;
1826 /* Adjust cursors pointing to mp */
1827 mc->mc_pg[mc->mc_top] = np;
1829 if (mc->mc_flags & C_SUB) {
1831 for (m2 = txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
1832 m3 = &m2->mc_xcursor->mx_cursor;
1833 if (m3->mc_snum < mc->mc_snum) continue;
1834 if (m3->mc_pg[mc->mc_top] == mp)
1835 m3->mc_pg[mc->mc_top] = np;
1838 for (m2 = txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
1839 if (m2->mc_snum < mc->mc_snum) continue;
1840 if (m2->mc_pg[mc->mc_top] == mp) {
1841 m2->mc_pg[mc->mc_top] = np;
1842 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
1843 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
1845 MDB_node *leaf = NODEPTR(np, mc->mc_ki[mc->mc_top]);
1846 if (!(leaf->mn_flags & F_SUBDATA))
1847 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
1856 mdb_env_sync(MDB_env *env, int force)
1859 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
1860 if (env->me_flags & MDB_WRITEMAP) {
1861 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
1862 ? MS_ASYNC : MS_SYNC;
1863 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
1866 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
1870 if (MDB_FDATASYNC(env->me_fd))
1877 /** Back up parent txn's cursors, then grab the originals for tracking */
1879 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
1881 MDB_cursor *mc, *bk;
1886 for (i = src->mt_numdbs; --i >= 0; ) {
1887 if ((mc = src->mt_cursors[i]) != NULL) {
1888 size = sizeof(MDB_cursor);
1890 size += sizeof(MDB_xcursor);
1891 for (; mc; mc = bk->mc_next) {
1897 mc->mc_db = &dst->mt_dbs[i];
1898 /* Kill pointers into src - and dst to reduce abuse: The
1899 * user may not use mc until dst ends. Otherwise we'd...
1901 mc->mc_txn = NULL; /* ...set this to dst */
1902 mc->mc_dbflag = NULL; /* ...and &dst->mt_dbflags[i] */
1903 if ((mx = mc->mc_xcursor) != NULL) {
1904 *(MDB_xcursor *)(bk+1) = *mx;
1905 mx->mx_cursor.mc_txn = NULL; /* ...and dst. */
1907 mc->mc_next = dst->mt_cursors[i];
1908 dst->mt_cursors[i] = mc;
1915 /** Close this write txn's cursors, give parent txn's cursors back to parent.
1916 * @param[in] txn the transaction handle.
1917 * @param[in] merge true to keep changes to parent cursors, false to revert.
1918 * @return 0 on success, non-zero on failure.
1921 mdb_cursors_close(MDB_txn *txn, unsigned merge)
1923 MDB_cursor **cursors = txn->mt_cursors, *mc, *next, *bk;
1927 for (i = txn->mt_numdbs; --i >= 0; ) {
1928 for (mc = cursors[i]; mc; mc = next) {
1930 if ((bk = mc->mc_backup) != NULL) {
1932 /* Commit changes to parent txn */
1933 mc->mc_next = bk->mc_next;
1934 mc->mc_backup = bk->mc_backup;
1935 mc->mc_txn = bk->mc_txn;
1936 mc->mc_db = bk->mc_db;
1937 mc->mc_dbflag = bk->mc_dbflag;
1938 if ((mx = mc->mc_xcursor) != NULL)
1939 mx->mx_cursor.mc_txn = bk->mc_txn;
1941 /* Abort nested txn */
1943 if ((mx = mc->mc_xcursor) != NULL)
1944 *mx = *(MDB_xcursor *)(bk+1);
1954 #ifdef MDB_DEBUG_SKIP
1955 #define mdb_txn_reset0(txn, act) mdb_txn_reset0(txn)
1958 mdb_txn_reset0(MDB_txn *txn, const char *act);
1960 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
1961 * @param[in] txn the transaction handle to initialize
1962 * @return 0 on success, non-zero on failure.
1965 mdb_txn_renew0(MDB_txn *txn)
1967 MDB_env *env = txn->mt_env;
1970 int rc, new_notls = 0;
1973 txn->mt_numdbs = env->me_numdbs;
1974 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
1976 if (txn->mt_flags & MDB_TXN_RDONLY) {
1977 if (!env->me_txns) {
1978 i = mdb_env_pick_meta(env);
1979 txn->mt_txnid = env->me_metas[i]->mm_txnid;
1980 txn->mt_u.reader = NULL;
1982 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
1983 pthread_getspecific(env->me_txkey);
1985 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
1986 return MDB_BAD_RSLOT;
1988 pid_t pid = env->me_pid;
1989 pthread_t tid = pthread_self();
1992 for (i=0; i<env->me_txns->mti_numreaders; i++)
1993 if (env->me_txns->mti_readers[i].mr_pid == 0)
1995 if (i == env->me_maxreaders) {
1996 UNLOCK_MUTEX_R(env);
1997 return MDB_READERS_FULL;
1999 env->me_txns->mti_readers[i].mr_pid = pid;
2000 env->me_txns->mti_readers[i].mr_tid = tid;
2001 if (i >= env->me_txns->mti_numreaders)
2002 env->me_txns->mti_numreaders = i+1;
2003 /* Save numreaders for un-mutexed mdb_env_close() */
2004 env->me_numreaders = env->me_txns->mti_numreaders;
2005 UNLOCK_MUTEX_R(env);
2006 r = &env->me_txns->mti_readers[i];
2007 new_notls = (env->me_flags & MDB_NOTLS);
2008 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
2013 txn->mt_txnid = r->mr_txnid = env->me_txns->mti_txnid;
2014 txn->mt_u.reader = r;
2016 txn->mt_toggle = txn->mt_txnid & 1;
2020 txn->mt_txnid = env->me_txns->mti_txnid;
2021 txn->mt_toggle = txn->mt_txnid & 1;
2024 if (txn->mt_txnid == mdb_debug_start)
2027 txn->mt_dirty_room = MDB_IDL_UM_MAX;
2028 txn->mt_u.dirty_list = env->me_dirty_list;
2029 txn->mt_u.dirty_list[0].mid = 0;
2030 txn->mt_free_pgs = env->me_free_pgs;
2031 txn->mt_free_pgs[0] = 0;
2032 txn->mt_spill_pgs = NULL;
2036 /* Copy the DB info and flags */
2037 memcpy(txn->mt_dbs, env->me_metas[txn->mt_toggle]->mm_dbs, 2 * sizeof(MDB_db));
2039 /* Moved to here to avoid a data race in read TXNs */
2040 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
2042 for (i=2; i<txn->mt_numdbs; i++) {
2043 x = env->me_dbflags[i];
2044 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
2045 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_STALE : 0;
2047 txn->mt_dbflags[0] = txn->mt_dbflags[1] = DB_VALID;
2049 if (env->me_maxpg < txn->mt_next_pgno) {
2050 mdb_txn_reset0(txn, "renew0-mapfail");
2052 txn->mt_u.reader->mr_pid = 0;
2053 txn->mt_u.reader = NULL;
2055 return MDB_MAP_RESIZED;
2062 mdb_txn_renew(MDB_txn *txn)
2066 if (!txn || txn->mt_dbxs) /* A reset txn has mt_dbxs==NULL */
2069 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
2070 DPUTS("environment had fatal error, must shutdown!");
2074 rc = mdb_txn_renew0(txn);
2075 if (rc == MDB_SUCCESS) {
2076 DPRINTF("renew txn %zu%c %p on mdbenv %p, root page %zu",
2077 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2078 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
2084 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
2088 int rc, size, tsize = sizeof(MDB_txn);
2090 if (env->me_flags & MDB_FATAL_ERROR) {
2091 DPUTS("environment had fatal error, must shutdown!");
2094 if ((env->me_flags & MDB_RDONLY) && !(flags & MDB_RDONLY))
2097 /* Nested transactions: Max 1 child, write txns only, no writemap */
2098 if (parent->mt_child ||
2099 (flags & MDB_RDONLY) || (parent->mt_flags & MDB_TXN_RDONLY) ||
2100 (env->me_flags & MDB_WRITEMAP))
2104 tsize = sizeof(MDB_ntxn);
2106 size = tsize + env->me_maxdbs * (sizeof(MDB_db)+1);
2107 if (!(flags & MDB_RDONLY))
2108 size += env->me_maxdbs * sizeof(MDB_cursor *);
2110 if ((txn = calloc(1, size)) == NULL) {
2111 DPRINTF("calloc: %s", strerror(ErrCode()));
2114 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
2115 if (flags & MDB_RDONLY) {
2116 txn->mt_flags |= MDB_TXN_RDONLY;
2117 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
2119 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
2120 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
2126 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
2127 if (!txn->mt_u.dirty_list ||
2128 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
2130 free(txn->mt_u.dirty_list);
2134 txn->mt_txnid = parent->mt_txnid;
2135 txn->mt_toggle = parent->mt_toggle;
2136 txn->mt_dirty_room = parent->mt_dirty_room;
2137 txn->mt_u.dirty_list[0].mid = 0;
2138 txn->mt_spill_pgs = NULL;
2139 txn->mt_next_pgno = parent->mt_next_pgno;
2140 parent->mt_child = txn;
2141 txn->mt_parent = parent;
2142 txn->mt_numdbs = parent->mt_numdbs;
2143 txn->mt_flags = parent->mt_flags;
2144 txn->mt_dbxs = parent->mt_dbxs;
2145 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2146 /* Copy parent's mt_dbflags, but clear DB_NEW */
2147 for (i=0; i<txn->mt_numdbs; i++)
2148 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
2150 ntxn = (MDB_ntxn *)txn;
2151 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
2152 if (env->me_pghead) {
2153 size = MDB_IDL_SIZEOF(env->me_pghead);
2154 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
2156 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
2161 rc = mdb_cursor_shadow(parent, txn);
2163 mdb_txn_reset0(txn, "beginchild-fail");
2165 rc = mdb_txn_renew0(txn);
2171 DPRINTF("begin txn %zu%c %p on mdbenv %p, root page %zu",
2172 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2173 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root);
2179 /** Export or close DBI handles opened in this txn. */
2181 mdb_dbis_update(MDB_txn *txn, int keep)
2184 MDB_dbi n = txn->mt_numdbs;
2185 MDB_env *env = txn->mt_env;
2186 unsigned char *tdbflags = txn->mt_dbflags;
2188 for (i = n; --i >= 2;) {
2189 if (tdbflags[i] & DB_NEW) {
2191 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2193 char *ptr = env->me_dbxs[i].md_name.mv_data;
2194 env->me_dbxs[i].md_name.mv_data = NULL;
2195 env->me_dbxs[i].md_name.mv_size = 0;
2196 env->me_dbflags[i] = 0;
2201 if (keep && env->me_numdbs < n)
2205 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
2206 * May be called twice for readonly txns: First reset it, then abort.
2207 * @param[in] txn the transaction handle to reset
2210 mdb_txn_reset0(MDB_txn *txn, const char *act)
2212 MDB_env *env = txn->mt_env;
2214 /* Close any DBI handles opened in this txn */
2215 mdb_dbis_update(txn, 0);
2217 DPRINTF("%s txn %zu%c %p on mdbenv %p, root page %zu",
2218 act, txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2219 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root);
2221 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2222 if (txn->mt_u.reader) {
2223 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2224 if (!(env->me_flags & MDB_NOTLS))
2225 txn->mt_u.reader = NULL; /* txn does not own reader */
2227 txn->mt_numdbs = 0; /* close nothing if called again */
2228 txn->mt_dbxs = NULL; /* mark txn as reset */
2230 mdb_cursors_close(txn, 0);
2232 if (!(env->me_flags & MDB_WRITEMAP)) {
2233 mdb_dlist_free(txn);
2235 mdb_midl_free(env->me_pghead);
2237 if (txn->mt_parent) {
2238 txn->mt_parent->mt_child = NULL;
2239 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2240 mdb_midl_free(txn->mt_free_pgs);
2241 mdb_midl_free(txn->mt_spill_pgs);
2242 free(txn->mt_u.dirty_list);
2246 if (mdb_midl_shrink(&txn->mt_free_pgs))
2247 env->me_free_pgs = txn->mt_free_pgs;
2248 env->me_pghead = NULL;
2252 /* The writer mutex was locked in mdb_txn_begin. */
2253 UNLOCK_MUTEX_W(env);
2258 mdb_txn_reset(MDB_txn *txn)
2263 /* This call is only valid for read-only txns */
2264 if (!(txn->mt_flags & MDB_TXN_RDONLY))
2267 mdb_txn_reset0(txn, "reset");
2271 mdb_txn_abort(MDB_txn *txn)
2277 mdb_txn_abort(txn->mt_child);
2279 mdb_txn_reset0(txn, "abort");
2280 /* Free reader slot tied to this txn (if MDB_NOTLS && writable FS) */
2281 if ((txn->mt_flags & MDB_TXN_RDONLY) && txn->mt_u.reader)
2282 txn->mt_u.reader->mr_pid = 0;
2287 /** Save the freelist as of this transaction to the freeDB.
2288 * This changes the freelist. Keep trying until it stabilizes.
2291 mdb_freelist_save(MDB_txn *txn)
2293 /* env->me_pghead[] can grow and shrink during this call.
2294 * env->me_pglast and txn->mt_free_pgs[] can only grow.
2295 * Page numbers cannot disappear from txn->mt_free_pgs[].
2298 MDB_env *env = txn->mt_env;
2299 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
2300 txnid_t pglast = 0, head_id = 0;
2301 pgno_t freecnt = 0, *free_pgs, *mop;
2302 ssize_t head_room = 0, total_room = 0, mop_len;
2304 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2306 if (env->me_pghead) {
2307 /* Make sure first page of freeDB is touched and on freelist */
2308 rc = mdb_page_search(&mc, NULL, MDB_PS_MODIFY);
2309 if (rc && rc != MDB_NOTFOUND)
2314 /* Come back here after each Put() in case freelist changed */
2317 /* If using records from freeDB which we have not yet
2318 * deleted, delete them and any we reserved for me_pghead.
2320 while (pglast < env->me_pglast) {
2321 rc = mdb_cursor_first(&mc, &key, NULL);
2324 pglast = head_id = *(txnid_t *)key.mv_data;
2325 total_room = head_room = 0;
2326 assert(pglast <= env->me_pglast);
2327 rc = mdb_cursor_del(&mc, 0);
2332 /* Save the IDL of pages freed by this txn, to a single record */
2333 if (freecnt < txn->mt_free_pgs[0]) {
2335 /* Make sure last page of freeDB is touched and on freelist */
2336 key.mv_size = MDB_MAXKEYSIZE+1;
2338 rc = mdb_page_search(&mc, &key, MDB_PS_MODIFY);
2339 if (rc && rc != MDB_NOTFOUND)
2342 free_pgs = txn->mt_free_pgs;
2343 /* Write to last page of freeDB */
2344 key.mv_size = sizeof(txn->mt_txnid);
2345 key.mv_data = &txn->mt_txnid;
2347 freecnt = free_pgs[0];
2348 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
2349 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2352 /* Retry if mt_free_pgs[] grew during the Put() */
2353 free_pgs = txn->mt_free_pgs;
2354 } while (freecnt < free_pgs[0]);
2355 mdb_midl_sort(free_pgs);
2356 memcpy(data.mv_data, free_pgs, data.mv_size);
2359 unsigned int i = free_pgs[0];
2360 DPRINTF("IDL write txn %zu root %zu num %u",
2361 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i);
2363 DPRINTF("IDL %zu", free_pgs[i]);
2369 mop = env->me_pghead;
2370 mop_len = mop ? mop[0] : 0;
2372 /* Reserve records for me_pghead[]. Split it if multi-page,
2373 * to avoid searching freeDB for a page range. Use keys in
2374 * range [1,me_pglast]: Smaller than txnid of oldest reader.
2376 if (total_room >= mop_len) {
2377 if (total_room == mop_len || --more < 0)
2379 } else if (head_room >= maxfree_1pg && head_id > 1) {
2380 /* Keep current record (overflow page), add a new one */
2384 /* (Re)write {key = head_id, IDL length = head_room} */
2385 total_room -= head_room;
2386 head_room = mop_len - total_room;
2387 if (head_room > maxfree_1pg && head_id > 1) {
2388 /* Overflow multi-page for part of me_pghead */
2389 head_room /= head_id; /* amortize page sizes */
2390 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
2391 } else if (head_room < 0) {
2392 /* Rare case, not bothering to delete this record */
2395 key.mv_size = sizeof(head_id);
2396 key.mv_data = &head_id;
2397 data.mv_size = (head_room + 1) * sizeof(pgno_t);
2398 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2401 *(MDB_ID *)data.mv_data = 0; /* IDL is initially empty */
2402 total_room += head_room;
2405 /* Fill in the reserved, touched me_pghead records */
2411 rc = mdb_cursor_first(&mc, &key, &data);
2412 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
2413 unsigned flags = MDB_CURRENT;
2414 txnid_t id = *(txnid_t *)key.mv_data;
2415 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
2418 assert(len >= 0 && id <= env->me_pglast);
2420 if (len > mop_len) {
2422 data.mv_size = (len + 1) * sizeof(MDB_ID);
2425 data.mv_data = mop -= len;
2428 rc = mdb_cursor_put(&mc, &key, &data, flags);
2430 if (rc || !(mop_len -= len))
2437 /** Flush dirty pages to the map, after clearing their dirty flag.
2440 mdb_page_flush(MDB_txn *txn)
2442 MDB_env *env = txn->mt_env;
2443 MDB_ID2L dl = txn->mt_u.dirty_list;
2444 unsigned psize = env->me_psize, j;
2445 int i, pagecount = dl[0].mid, rc;
2446 size_t size = 0, pos = 0;
2448 MDB_page *dp = NULL;
2452 struct iovec iov[MDB_COMMIT_PAGES];
2453 ssize_t wpos = 0, wsize = 0, wres;
2454 size_t next_pos = 1; /* impossible pos, so pos != next_pos */
2459 if (env->me_flags & MDB_WRITEMAP) {
2460 /* Clear dirty flags */
2461 for (i = pagecount; i; i--) {
2463 /* Don't flush this page yet */
2464 if (dp->mp_flags & P_KEEP) {
2465 dp->mp_flags ^= P_KEEP;
2469 dp->mp_flags &= ~P_DIRTY;
2475 /* Write the pages */
2477 if (i <= pagecount) {
2479 /* Don't flush this page yet */
2480 if (dp->mp_flags & P_KEEP) {
2481 dp->mp_flags ^= P_KEEP;
2486 /* clear dirty flag */
2487 dp->mp_flags &= ~P_DIRTY;
2490 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
2495 /* Windows actually supports scatter/gather I/O, but only on
2496 * unbuffered file handles. Since we're relying on the OS page
2497 * cache for all our data, that's self-defeating. So we just
2498 * write pages one at a time. We use the ov structure to set
2499 * the write offset, to at least save the overhead of a Seek
2502 DPRINTF("committing page %zu", pgno);
2503 memset(&ov, 0, sizeof(ov));
2504 ov.Offset = pos & 0xffffffff;
2505 ov.OffsetHigh = pos >> 16 >> 16;
2506 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
2508 DPRINTF("WriteFile: %d", rc);
2512 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
2513 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
2515 /* Write previous page(s) */
2516 #ifdef MDB_USE_PWRITEV
2517 wres = pwritev(env->me_fd, iov, n, wpos);
2520 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
2522 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
2524 DPRINTF("lseek: %s", strerror(rc));
2527 wres = writev(env->me_fd, iov, n);
2530 if (wres != wsize) {
2533 DPRINTF("Write error: %s", strerror(rc));
2535 rc = EIO; /* TODO: Use which error code? */
2536 DPUTS("short write, filesystem full?");
2547 DPRINTF("committing page %zu", pgno);
2548 next_pos = pos + size;
2549 iov[n].iov_len = size;
2550 iov[n].iov_base = (char *)dp;
2557 for (i=1; i<=pagecount; i++) {
2559 /* This is a page we skipped above */
2562 dl[j].mid = dp->mp_pgno;
2565 mdb_dpage_free(env, dp);
2573 mdb_txn_commit(MDB_txn *txn)
2579 assert(txn != NULL);
2580 assert(txn->mt_env != NULL);
2582 if (txn->mt_child) {
2583 rc = mdb_txn_commit(txn->mt_child);
2584 txn->mt_child = NULL;
2591 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2592 mdb_dbis_update(txn, 1);
2593 txn->mt_numdbs = 2; /* so txn_abort() doesn't close any new handles */
2598 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
2599 DPUTS("error flag is set, can't commit");
2601 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
2606 if (txn->mt_parent) {
2607 MDB_txn *parent = txn->mt_parent;
2611 /* Append our free list to parent's */
2612 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
2615 mdb_midl_free(txn->mt_free_pgs);
2617 parent->mt_next_pgno = txn->mt_next_pgno;
2618 parent->mt_flags = txn->mt_flags;
2620 /* Merge our cursors into parent's and close them */
2621 mdb_cursors_close(txn, 1);
2623 /* Update parent's DB table. */
2624 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2625 parent->mt_numdbs = txn->mt_numdbs;
2626 parent->mt_dbflags[0] = txn->mt_dbflags[0];
2627 parent->mt_dbflags[1] = txn->mt_dbflags[1];
2628 for (i=2; i<txn->mt_numdbs; i++) {
2629 /* preserve parent's DB_NEW status */
2630 x = parent->mt_dbflags[i] & DB_NEW;
2631 parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
2634 dst = parent->mt_u.dirty_list;
2635 src = txn->mt_u.dirty_list;
2636 /* Remove anything in our dirty list from parent's spill list */
2637 if (parent->mt_spill_pgs) {
2638 x = parent->mt_spill_pgs[0];
2640 /* zero out our dirty pages in parent spill list */
2641 for (i=1; i<=src[0].mid; i++) {
2642 if (src[i].mid < parent->mt_spill_pgs[x])
2644 if (src[i].mid > parent->mt_spill_pgs[x]) {
2650 parent->mt_spill_pgs[x] = 0;
2653 /* OK, we had a few hits, squash zeros from the spill list */
2654 if (len < parent->mt_spill_pgs[0]) {
2656 for (y=1; y<=parent->mt_spill_pgs[0]; y++) {
2657 if (parent->mt_spill_pgs[y]) {
2659 parent->mt_spill_pgs[x] = parent->mt_spill_pgs[y];
2664 parent->mt_spill_pgs[0] = len;
2667 /* Find len = length of merging our dirty list with parent's */
2669 dst[0].mid = 0; /* simplify loops */
2670 if (parent->mt_parent) {
2671 len = x + src[0].mid;
2672 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
2673 for (i = x; y && i; y--) {
2674 pgno_t yp = src[y].mid;
2675 while (yp < dst[i].mid)
2677 if (yp == dst[i].mid) {
2682 } else { /* Simplify the above for single-ancestor case */
2683 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
2685 /* Merge our dirty list with parent's */
2687 for (i = len; y; dst[i--] = src[y--]) {
2688 pgno_t yp = src[y].mid;
2689 while (yp < dst[x].mid)
2690 dst[i--] = dst[x--];
2691 if (yp == dst[x].mid)
2692 free(dst[x--].mptr);
2696 free(txn->mt_u.dirty_list);
2697 parent->mt_dirty_room = txn->mt_dirty_room;
2698 if (txn->mt_spill_pgs) {
2699 if (parent->mt_spill_pgs) {
2700 mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs);
2701 mdb_midl_free(txn->mt_spill_pgs);
2702 mdb_midl_sort(parent->mt_spill_pgs);
2704 parent->mt_spill_pgs = txn->mt_spill_pgs;
2708 parent->mt_child = NULL;
2709 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
2714 if (txn != env->me_txn) {
2715 DPUTS("attempt to commit unknown transaction");
2720 mdb_cursors_close(txn, 0);
2722 if (!txn->mt_u.dirty_list[0].mid && !(txn->mt_flags & MDB_TXN_DIRTY))
2725 DPRINTF("committing txn %zu %p on mdbenv %p, root page %zu",
2726 txn->mt_txnid, (void *)txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root);
2728 /* Update DB root pointers */
2729 if (txn->mt_numdbs > 2) {
2733 data.mv_size = sizeof(MDB_db);
2735 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
2736 for (i = 2; i < txn->mt_numdbs; i++) {
2737 if (txn->mt_dbflags[i] & DB_DIRTY) {
2738 data.mv_data = &txn->mt_dbs[i];
2739 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
2746 rc = mdb_freelist_save(txn);
2750 mdb_midl_free(env->me_pghead);
2751 env->me_pghead = NULL;
2752 if (mdb_midl_shrink(&txn->mt_free_pgs))
2753 env->me_free_pgs = txn->mt_free_pgs;
2759 if ((rc = mdb_page_flush(txn)) ||
2760 (rc = mdb_env_sync(env, 0)) ||
2761 (rc = mdb_env_write_meta(txn)))
2767 mdb_dbis_update(txn, 1);
2769 UNLOCK_MUTEX_W(env);
2779 /** Read the environment parameters of a DB environment before
2780 * mapping it into memory.
2781 * @param[in] env the environment handle
2782 * @param[out] meta address of where to store the meta information
2783 * @return 0 on success, non-zero on failure.
2786 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
2793 /* We don't know the page size yet, so use a minimum value.
2794 * Read both meta pages so we can use the latest one.
2797 for (i=off=0; i<2; i++, off = meta->mm_psize) {
2801 memset(&ov, 0, sizeof(ov));
2803 rc = ReadFile(env->me_fd,&pbuf,MDB_PAGESIZE,&len,&ov) ? (int)len : -1;
2804 if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF)
2807 rc = pread(env->me_fd, &pbuf, MDB_PAGESIZE, off);
2809 if (rc != MDB_PAGESIZE) {
2810 if (rc == 0 && off == 0)
2812 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
2813 DPRINTF("read: %s", mdb_strerror(rc));
2817 p = (MDB_page *)&pbuf;
2819 if (!F_ISSET(p->mp_flags, P_META)) {
2820 DPRINTF("page %zu not a meta page", p->mp_pgno);
2825 if (m->mm_magic != MDB_MAGIC) {
2826 DPUTS("meta has invalid magic");
2830 if (m->mm_version != MDB_VERSION) {
2831 DPRINTF("database is version %u, expected version %u",
2832 m->mm_version, MDB_VERSION);
2833 return MDB_VERSION_MISMATCH;
2836 if (off == 0 || m->mm_txnid > meta->mm_txnid)
2842 /** Write the environment parameters of a freshly created DB environment.
2843 * @param[in] env the environment handle
2844 * @param[out] meta address of where to store the meta information
2845 * @return 0 on success, non-zero on failure.
2848 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
2854 DPUTS("writing new meta page");
2856 GET_PAGESIZE(psize);
2858 meta->mm_magic = MDB_MAGIC;
2859 meta->mm_version = MDB_VERSION;
2860 meta->mm_mapsize = env->me_mapsize;
2861 meta->mm_psize = psize;
2862 meta->mm_last_pg = 1;
2863 meta->mm_flags = env->me_flags & 0xffff;
2864 meta->mm_flags |= MDB_INTEGERKEY;
2865 meta->mm_dbs[0].md_root = P_INVALID;
2866 meta->mm_dbs[1].md_root = P_INVALID;
2868 p = calloc(2, psize);
2870 p->mp_flags = P_META;
2871 *(MDB_meta *)METADATA(p) = *meta;
2873 q = (MDB_page *)((char *)p + psize);
2875 q->mp_flags = P_META;
2876 *(MDB_meta *)METADATA(q) = *meta;
2882 memset(&ov, 0, sizeof(ov));
2883 rc = WriteFile(env->me_fd, p, psize * 2, &len, &ov);
2884 rc = rc ? (len == psize * 2 ? MDB_SUCCESS : EIO) : ErrCode();
2887 rc = pwrite(env->me_fd, p, psize * 2, 0);
2888 rc = (rc == (int)psize * 2) ? MDB_SUCCESS : rc < 0 ? ErrCode() : EIO;
2894 /** Update the environment info to commit a transaction.
2895 * @param[in] txn the transaction that's being committed
2896 * @return 0 on success, non-zero on failure.
2899 mdb_env_write_meta(MDB_txn *txn)
2902 MDB_meta meta, metab, *mp;
2904 int rc, len, toggle;
2913 assert(txn != NULL);
2914 assert(txn->mt_env != NULL);
2916 toggle = !txn->mt_toggle;
2917 DPRINTF("writing meta page %d for root page %zu",
2918 toggle, txn->mt_dbs[MAIN_DBI].md_root);
2921 mp = env->me_metas[toggle];
2923 if (env->me_flags & MDB_WRITEMAP) {
2924 /* Persist any increases of mapsize config */
2925 if (env->me_mapsize > mp->mm_mapsize)
2926 mp->mm_mapsize = env->me_mapsize;
2927 mp->mm_dbs[0] = txn->mt_dbs[0];
2928 mp->mm_dbs[1] = txn->mt_dbs[1];
2929 mp->mm_last_pg = txn->mt_next_pgno - 1;
2930 mp->mm_txnid = txn->mt_txnid;
2931 if (!(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
2932 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
2935 ptr += env->me_psize;
2936 if (MDB_MSYNC(ptr, env->me_psize, rc)) {
2943 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
2944 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
2946 ptr = (char *)&meta;
2947 if (env->me_mapsize > mp->mm_mapsize) {
2948 /* Persist any increases of mapsize config */
2949 meta.mm_mapsize = env->me_mapsize;
2950 off = offsetof(MDB_meta, mm_mapsize);
2952 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
2954 len = sizeof(MDB_meta) - off;
2957 meta.mm_dbs[0] = txn->mt_dbs[0];
2958 meta.mm_dbs[1] = txn->mt_dbs[1];
2959 meta.mm_last_pg = txn->mt_next_pgno - 1;
2960 meta.mm_txnid = txn->mt_txnid;
2963 off += env->me_psize;
2966 /* Write to the SYNC fd */
2967 mfd = env->me_flags & (MDB_NOSYNC|MDB_NOMETASYNC) ?
2968 env->me_fd : env->me_mfd;
2971 memset(&ov, 0, sizeof(ov));
2973 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
2977 rc = pwrite(mfd, ptr, len, off);
2980 rc = rc < 0 ? ErrCode() : EIO;
2981 DPUTS("write failed, disk error?");
2982 /* On a failure, the pagecache still contains the new data.
2983 * Write some old data back, to prevent it from being used.
2984 * Use the non-SYNC fd; we know it will fail anyway.
2986 meta.mm_last_pg = metab.mm_last_pg;
2987 meta.mm_txnid = metab.mm_txnid;
2989 memset(&ov, 0, sizeof(ov));
2991 WriteFile(env->me_fd, ptr, len, NULL, &ov);
2993 r2 = pwrite(env->me_fd, ptr, len, off);
2996 env->me_flags |= MDB_FATAL_ERROR;
3000 /* Memory ordering issues are irrelevant; since the entire writer
3001 * is wrapped by wmutex, all of these changes will become visible
3002 * after the wmutex is unlocked. Since the DB is multi-version,
3003 * readers will get consistent data regardless of how fresh or
3004 * how stale their view of these values is.
3006 env->me_txns->mti_txnid = txn->mt_txnid;
3011 /** Check both meta pages to see which one is newer.
3012 * @param[in] env the environment handle
3013 * @return meta toggle (0 or 1).
3016 mdb_env_pick_meta(const MDB_env *env)
3018 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
3022 mdb_env_create(MDB_env **env)
3026 e = calloc(1, sizeof(MDB_env));
3030 e->me_maxreaders = DEFAULT_READERS;
3031 e->me_maxdbs = e->me_numdbs = 2;
3032 e->me_fd = INVALID_HANDLE_VALUE;
3033 e->me_lfd = INVALID_HANDLE_VALUE;
3034 e->me_mfd = INVALID_HANDLE_VALUE;
3035 #ifdef MDB_USE_POSIX_SEM
3036 e->me_rmutex = SEM_FAILED;
3037 e->me_wmutex = SEM_FAILED;
3039 e->me_pid = getpid();
3040 VGMEMP_CREATE(e,0,0);
3046 mdb_env_set_mapsize(MDB_env *env, size_t size)
3050 env->me_mapsize = size;
3052 env->me_maxpg = env->me_mapsize / env->me_psize;
3057 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
3061 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
3066 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
3068 if (env->me_map || readers < 1)
3070 env->me_maxreaders = readers;
3075 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
3077 if (!env || !readers)
3079 *readers = env->me_maxreaders;
3083 /** Further setup required for opening an MDB environment
3086 mdb_env_open2(MDB_env *env)
3088 unsigned int flags = env->me_flags;
3096 memset(&meta, 0, sizeof(meta));
3098 if ((i = mdb_env_read_header(env, &meta)) != 0) {
3101 DPUTS("new mdbenv");
3105 /* Was a mapsize configured? */
3106 if (!env->me_mapsize) {
3107 /* If this is a new environment, take the default,
3108 * else use the size recorded in the existing env.
3110 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
3111 } else if (env->me_mapsize < meta.mm_mapsize) {
3112 /* If the configured size is smaller, make sure it's
3113 * still big enough. Silently round up to minimum if not.
3115 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
3116 if (env->me_mapsize < minsize)
3117 env->me_mapsize = minsize;
3124 LONG sizelo, sizehi;
3125 sizelo = env->me_mapsize & 0xffffffff;
3126 sizehi = env->me_mapsize >> 16 >> 16; /* only needed on Win64 */
3127 /* Windows won't create mappings for zero length files.
3128 * Just allocate the maxsize right now.
3131 if (SetFilePointer(env->me_fd, sizelo, &sizehi, 0) != (DWORD)sizelo
3132 || !SetEndOfFile(env->me_fd)
3133 || SetFilePointer(env->me_fd, 0, NULL, 0) != 0)
3136 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
3137 PAGE_READWRITE : PAGE_READONLY,
3138 sizehi, sizelo, NULL);
3141 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
3142 FILE_MAP_WRITE : FILE_MAP_READ,
3143 0, 0, env->me_mapsize, meta.mm_address);
3144 rc = env->me_map ? 0 : ErrCode();
3152 if (flags & MDB_WRITEMAP) {
3154 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
3157 env->me_map = mmap(meta.mm_address, env->me_mapsize, prot, i,
3159 if (env->me_map == MAP_FAILED) {
3163 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
3165 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
3167 #ifdef POSIX_MADV_RANDOM
3168 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
3169 #endif /* POSIX_MADV_RANDOM */
3170 #endif /* MADV_RANDOM */
3174 if (flags & MDB_FIXEDMAP)
3175 meta.mm_address = env->me_map;
3176 i = mdb_env_init_meta(env, &meta);
3177 if (i != MDB_SUCCESS) {
3180 } else if (meta.mm_address && env->me_map != meta.mm_address) {
3181 /* Can happen because the address argument to mmap() is just a
3182 * hint. mmap() can pick another, e.g. if the range is in use.
3183 * The MAP_FIXED flag would prevent that, but then mmap could
3184 * instead unmap existing pages to make room for the new map.
3186 return EBUSY; /* TODO: Make a new MDB_* error code? */
3188 env->me_psize = meta.mm_psize;
3189 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
3190 env->me_nodemax = (env->me_psize - PAGEHDRSZ) / MDB_MINKEYS;
3192 env->me_maxpg = env->me_mapsize / env->me_psize;
3194 p = (MDB_page *)env->me_map;
3195 env->me_metas[0] = METADATA(p);
3196 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + meta.mm_psize);
3200 int toggle = mdb_env_pick_meta(env);
3201 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
3203 DPRINTF("opened database version %u, pagesize %u",
3204 env->me_metas[0]->mm_version, env->me_psize);
3205 DPRINTF("using meta page %d", toggle);
3206 DPRINTF("depth: %u", db->md_depth);
3207 DPRINTF("entries: %zu", db->md_entries);
3208 DPRINTF("branch pages: %zu", db->md_branch_pages);
3209 DPRINTF("leaf pages: %zu", db->md_leaf_pages);
3210 DPRINTF("overflow pages: %zu", db->md_overflow_pages);
3211 DPRINTF("root: %zu", db->md_root);
3219 /** Release a reader thread's slot in the reader lock table.
3220 * This function is called automatically when a thread exits.
3221 * @param[in] ptr This points to the slot in the reader lock table.
3224 mdb_env_reader_dest(void *ptr)
3226 MDB_reader *reader = ptr;
3232 /** Junk for arranging thread-specific callbacks on Windows. This is
3233 * necessarily platform and compiler-specific. Windows supports up
3234 * to 1088 keys. Let's assume nobody opens more than 64 environments
3235 * in a single process, for now. They can override this if needed.
3237 #ifndef MAX_TLS_KEYS
3238 #define MAX_TLS_KEYS 64
3240 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
3241 static int mdb_tls_nkeys;
3243 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
3247 case DLL_PROCESS_ATTACH: break;
3248 case DLL_THREAD_ATTACH: break;
3249 case DLL_THREAD_DETACH:
3250 for (i=0; i<mdb_tls_nkeys; i++) {
3251 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
3252 mdb_env_reader_dest(r);
3255 case DLL_PROCESS_DETACH: break;
3260 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3262 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3266 /* Force some symbol references.
3267 * _tls_used forces the linker to create the TLS directory if not already done
3268 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
3270 #pragma comment(linker, "/INCLUDE:_tls_used")
3271 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
3272 #pragma const_seg(".CRT$XLB")
3273 extern const PIMAGE_TLS_CALLBACK mdb_tls_callback;
3274 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3277 #pragma comment(linker, "/INCLUDE:__tls_used")
3278 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
3279 #pragma data_seg(".CRT$XLB")
3280 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3282 #endif /* WIN 32/64 */
3283 #endif /* !__GNUC__ */
3286 /** Downgrade the exclusive lock on the region back to shared */
3288 mdb_env_share_locks(MDB_env *env, int *excl)
3290 int rc = 0, toggle = mdb_env_pick_meta(env);
3292 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
3297 /* First acquire a shared lock. The Unlock will
3298 * then release the existing exclusive lock.
3300 memset(&ov, 0, sizeof(ov));
3301 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3304 UnlockFile(env->me_lfd, 0, 0, 1, 0);
3310 struct flock lock_info;
3311 /* The shared lock replaces the existing lock */
3312 memset((void *)&lock_info, 0, sizeof(lock_info));
3313 lock_info.l_type = F_RDLCK;
3314 lock_info.l_whence = SEEK_SET;
3315 lock_info.l_start = 0;
3316 lock_info.l_len = 1;
3317 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3318 (rc = ErrCode()) == EINTR) ;
3319 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
3326 /** Try to get exlusive lock, otherwise shared.
3327 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
3330 mdb_env_excl_lock(MDB_env *env, int *excl)
3334 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
3338 memset(&ov, 0, sizeof(ov));
3339 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3346 struct flock lock_info;
3347 memset((void *)&lock_info, 0, sizeof(lock_info));
3348 lock_info.l_type = F_WRLCK;
3349 lock_info.l_whence = SEEK_SET;
3350 lock_info.l_start = 0;
3351 lock_info.l_len = 1;
3352 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3353 (rc = ErrCode()) == EINTR) ;
3357 # ifdef MDB_USE_POSIX_SEM
3358 if (*excl < 0) /* always true when !MDB_USE_POSIX_SEM */
3361 lock_info.l_type = F_RDLCK;
3362 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
3363 (rc = ErrCode()) == EINTR) ;
3371 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
3373 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
3375 * @(#) $Revision: 5.1 $
3376 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
3377 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
3379 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
3383 * Please do not copyright this code. This code is in the public domain.
3385 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
3386 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
3387 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
3388 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
3389 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
3390 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
3391 * PERFORMANCE OF THIS SOFTWARE.
3394 * chongo <Landon Curt Noll> /\oo/\
3395 * http://www.isthe.com/chongo/
3397 * Share and Enjoy! :-)
3400 typedef unsigned long long mdb_hash_t;
3401 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
3403 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
3404 * @param[in] str string to hash
3405 * @param[in] hval initial value for hash
3406 * @return 64 bit hash
3408 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
3409 * hval arg on the first call.
3412 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
3414 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
3415 unsigned char *end = s + val->mv_size;
3417 * FNV-1a hash each octet of the string
3420 /* xor the bottom with the current octet */
3421 hval ^= (mdb_hash_t)*s++;
3423 /* multiply by the 64 bit FNV magic prime mod 2^64 */
3424 hval += (hval << 1) + (hval << 4) + (hval << 5) +
3425 (hval << 7) + (hval << 8) + (hval << 40);
3427 /* return our new hash value */
3431 /** Hash the string and output the hash in hex.
3432 * @param[in] str string to hash
3433 * @param[out] hexbuf an array of 17 chars to hold the hash
3436 mdb_hash_hex(MDB_val *val, char *hexbuf)
3439 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
3440 for (i=0; i<8; i++) {
3441 hexbuf += sprintf(hexbuf, "%02x", (unsigned int)h & 0xff);
3447 /** Open and/or initialize the lock region for the environment.
3448 * @param[in] env The MDB environment.
3449 * @param[in] lpath The pathname of the file used for the lock region.
3450 * @param[in] mode The Unix permissions for the file, if we create it.
3451 * @param[out] excl Resulting file lock type: -1 none, 0 shared, 1 exclusive
3452 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
3453 * @return 0 on success, non-zero on failure.
3456 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
3459 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
3461 # define MDB_ERRCODE_ROFS EROFS
3462 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
3463 # define MDB_CLOEXEC O_CLOEXEC
3466 # define MDB_CLOEXEC 0
3473 env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
3474 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
3475 FILE_ATTRIBUTE_NORMAL, NULL);
3477 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
3479 if (env->me_lfd == INVALID_HANDLE_VALUE) {
3481 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
3486 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
3487 /* Lose record locks when exec*() */
3488 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
3489 fcntl(env->me_lfd, F_SETFD, fdflags);
3492 if (!(env->me_flags & MDB_NOTLS)) {
3493 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
3496 env->me_flags |= MDB_ENV_TXKEY;
3498 /* Windows TLS callbacks need help finding their TLS info. */
3499 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
3503 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
3507 /* Try to get exclusive lock. If we succeed, then
3508 * nobody is using the lock region and we should initialize it.
3510 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
3513 size = GetFileSize(env->me_lfd, NULL);
3515 size = lseek(env->me_lfd, 0, SEEK_END);
3516 if (size == -1) goto fail_errno;
3518 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
3519 if (size < rsize && *excl > 0) {
3521 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != rsize
3522 || !SetEndOfFile(env->me_lfd))
3525 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
3529 size = rsize - sizeof(MDB_txninfo);
3530 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
3535 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
3537 if (!mh) goto fail_errno;
3538 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
3540 if (!env->me_txns) goto fail_errno;
3542 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
3544 if (m == MAP_FAILED) goto fail_errno;
3550 BY_HANDLE_FILE_INFORMATION stbuf;
3559 if (!mdb_sec_inited) {
3560 InitializeSecurityDescriptor(&mdb_null_sd,
3561 SECURITY_DESCRIPTOR_REVISION);
3562 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
3563 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
3564 mdb_all_sa.bInheritHandle = FALSE;
3565 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
3568 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
3569 idbuf.volume = stbuf.dwVolumeSerialNumber;
3570 idbuf.nhigh = stbuf.nFileIndexHigh;
3571 idbuf.nlow = stbuf.nFileIndexLow;
3572 val.mv_data = &idbuf;
3573 val.mv_size = sizeof(idbuf);
3574 mdb_hash_hex(&val, hexbuf);
3575 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", hexbuf);
3576 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", hexbuf);
3577 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
3578 if (!env->me_rmutex) goto fail_errno;
3579 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
3580 if (!env->me_wmutex) goto fail_errno;
3581 #elif defined(MDB_USE_POSIX_SEM)
3590 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
3591 idbuf.dev = stbuf.st_dev;
3592 idbuf.ino = stbuf.st_ino;
3593 val.mv_data = &idbuf;
3594 val.mv_size = sizeof(idbuf);
3595 mdb_hash_hex(&val, hexbuf);
3596 sprintf(env->me_txns->mti_rmname, "/MDBr%s", hexbuf);
3597 sprintf(env->me_txns->mti_wmname, "/MDBw%s", hexbuf);
3598 /* Clean up after a previous run, if needed: Try to
3599 * remove both semaphores before doing anything else.
3601 sem_unlink(env->me_txns->mti_rmname);
3602 sem_unlink(env->me_txns->mti_wmname);
3603 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
3604 O_CREAT|O_EXCL, mode, 1);
3605 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3606 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
3607 O_CREAT|O_EXCL, mode, 1);
3608 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3609 #else /* MDB_USE_POSIX_SEM */
3610 pthread_mutexattr_t mattr;
3612 if ((rc = pthread_mutexattr_init(&mattr))
3613 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
3614 || (rc = pthread_mutex_init(&env->me_txns->mti_mutex, &mattr))
3615 || (rc = pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr)))
3617 pthread_mutexattr_destroy(&mattr);
3618 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
3620 env->me_txns->mti_version = MDB_VERSION;
3621 env->me_txns->mti_magic = MDB_MAGIC;
3622 env->me_txns->mti_txnid = 0;
3623 env->me_txns->mti_numreaders = 0;
3626 if (env->me_txns->mti_magic != MDB_MAGIC) {
3627 DPUTS("lock region has invalid magic");
3631 if (env->me_txns->mti_version != MDB_VERSION) {
3632 DPRINTF("lock region is version %u, expected version %u",
3633 env->me_txns->mti_version, MDB_VERSION);
3634 rc = MDB_VERSION_MISMATCH;
3638 if (rc && rc != EACCES && rc != EAGAIN) {
3642 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
3643 if (!env->me_rmutex) goto fail_errno;
3644 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
3645 if (!env->me_wmutex) goto fail_errno;
3646 #elif defined(MDB_USE_POSIX_SEM)
3647 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
3648 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3649 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
3650 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3661 /** The name of the lock file in the DB environment */
3662 #define LOCKNAME "/lock.mdb"
3663 /** The name of the data file in the DB environment */
3664 #define DATANAME "/data.mdb"
3665 /** The suffix of the lock file when no subdir is used */
3666 #define LOCKSUFF "-lock"
3667 /** Only a subset of the @ref mdb_env flags can be changed
3668 * at runtime. Changing other flags requires closing the
3669 * environment and re-opening it with the new flags.
3671 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC)
3672 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY|MDB_WRITEMAP|MDB_NOTLS)
3675 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
3677 int oflags, rc, len, excl = -1;
3678 char *lpath, *dpath;
3680 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
3684 if (flags & MDB_NOSUBDIR) {
3685 rc = len + sizeof(LOCKSUFF) + len + 1;
3687 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
3692 if (flags & MDB_NOSUBDIR) {
3693 dpath = lpath + len + sizeof(LOCKSUFF);
3694 sprintf(lpath, "%s" LOCKSUFF, path);
3695 strcpy(dpath, path);
3697 dpath = lpath + len + sizeof(LOCKNAME);
3698 sprintf(lpath, "%s" LOCKNAME, path);
3699 sprintf(dpath, "%s" DATANAME, path);
3703 flags |= env->me_flags;
3704 if (flags & MDB_RDONLY) {
3705 /* silently ignore WRITEMAP when we're only getting read access */
3706 flags &= ~MDB_WRITEMAP;
3708 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
3709 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
3712 env->me_flags = flags |= MDB_ENV_ACTIVE;
3716 env->me_path = strdup(path);
3717 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
3718 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
3719 if (!(env->me_dbxs && env->me_path && env->me_dbflags)) {
3724 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
3729 if (F_ISSET(flags, MDB_RDONLY)) {
3730 oflags = GENERIC_READ;
3731 len = OPEN_EXISTING;
3733 oflags = GENERIC_READ|GENERIC_WRITE;
3736 mode = FILE_ATTRIBUTE_NORMAL;
3737 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
3738 NULL, len, mode, NULL);
3740 if (F_ISSET(flags, MDB_RDONLY))
3743 oflags = O_RDWR | O_CREAT;
3745 env->me_fd = open(dpath, oflags, mode);
3747 if (env->me_fd == INVALID_HANDLE_VALUE) {
3752 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
3753 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
3754 env->me_mfd = env->me_fd;
3756 /* Synchronous fd for meta writes. Needed even with
3757 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
3760 env->me_mfd = CreateFile(dpath, oflags,
3761 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
3762 mode | FILE_FLAG_WRITE_THROUGH, NULL);
3764 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
3766 if (env->me_mfd == INVALID_HANDLE_VALUE) {
3771 DPRINTF("opened dbenv %p", (void *) env);
3773 rc = mdb_env_share_locks(env, &excl);
3779 mdb_env_close0(env, excl);
3785 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
3787 mdb_env_close0(MDB_env *env, int excl)
3791 if (!(env->me_flags & MDB_ENV_ACTIVE))
3794 /* Doing this here since me_dbxs may not exist during mdb_env_close */
3795 for (i = env->me_maxdbs; --i > MAIN_DBI; )
3796 free(env->me_dbxs[i].md_name.mv_data);
3798 free(env->me_dbflags);
3801 free(env->me_dirty_list);
3802 mdb_midl_free(env->me_free_pgs);
3804 if (env->me_flags & MDB_ENV_TXKEY) {
3805 pthread_key_delete(env->me_txkey);
3807 /* Delete our key from the global list */
3808 for (i=0; i<mdb_tls_nkeys; i++)
3809 if (mdb_tls_keys[i] == env->me_txkey) {
3810 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
3818 munmap(env->me_map, env->me_mapsize);
3820 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
3821 (void) close(env->me_mfd);
3822 if (env->me_fd != INVALID_HANDLE_VALUE)
3823 (void) close(env->me_fd);
3825 pid_t pid = env->me_pid;
3826 /* Clearing readers is done in this function because
3827 * me_txkey with its destructor must be disabled first.
3829 for (i = env->me_numreaders; --i >= 0; )
3830 if (env->me_txns->mti_readers[i].mr_pid == pid)
3831 env->me_txns->mti_readers[i].mr_pid = 0;
3833 if (env->me_rmutex) {
3834 CloseHandle(env->me_rmutex);
3835 if (env->me_wmutex) CloseHandle(env->me_wmutex);
3837 /* Windows automatically destroys the mutexes when
3838 * the last handle closes.
3840 #elif defined(MDB_USE_POSIX_SEM)
3841 if (env->me_rmutex != SEM_FAILED) {
3842 sem_close(env->me_rmutex);
3843 if (env->me_wmutex != SEM_FAILED)
3844 sem_close(env->me_wmutex);
3845 /* If we have the filelock: If we are the
3846 * only remaining user, clean up semaphores.
3849 mdb_env_excl_lock(env, &excl);
3851 sem_unlink(env->me_txns->mti_rmname);
3852 sem_unlink(env->me_txns->mti_wmname);
3856 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
3858 if (env->me_lfd != INVALID_HANDLE_VALUE) {
3861 /* Unlock the lockfile. Windows would have unlocked it
3862 * after closing anyway, but not necessarily at once.
3864 UnlockFile(env->me_lfd, 0, 0, 1, 0);
3867 (void) close(env->me_lfd);
3870 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
3874 mdb_env_copyfd(MDB_env *env, HANDLE fd)
3876 MDB_txn *txn = NULL;
3881 /* Do the lock/unlock of the reader mutex before starting the
3882 * write txn. Otherwise other read txns could block writers.
3884 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
3889 /* We must start the actual read txn after blocking writers */
3890 mdb_txn_reset0(txn, "reset-stage1");
3892 /* Temporarily block writers until we snapshot the meta pages */
3895 rc = mdb_txn_renew0(txn);
3897 UNLOCK_MUTEX_W(env);
3902 wsize = env->me_psize * 2;
3906 rc = WriteFile(fd, env->me_map, wsize, &len, NULL);
3907 rc = rc ? (len == wsize ? MDB_SUCCESS : EIO) : ErrCode();
3910 rc = write(fd, env->me_map, wsize);
3911 rc = rc == (int)wsize ? MDB_SUCCESS : rc < 0 ? ErrCode() : EIO;
3914 UNLOCK_MUTEX_W(env);
3919 ptr = env->me_map + wsize;
3920 wsize = txn->mt_next_pgno * env->me_psize - wsize;
3924 if (wsize > MAX_WRITE)
3928 rc = WriteFile(fd, ptr, w2, &len, NULL);
3929 rc = rc ? (len == w2 ? MDB_SUCCESS : EIO) : ErrCode();
3938 if (wsize > MAX_WRITE)
3942 wres = write(fd, ptr, w2);
3943 rc = wres == (ssize_t)w2 ? MDB_SUCCESS : wres < 0 ? ErrCode() : EIO;
3956 mdb_env_copy(MDB_env *env, const char *path)
3960 HANDLE newfd = INVALID_HANDLE_VALUE;
3962 if (env->me_flags & MDB_NOSUBDIR) {
3963 lpath = (char *)path;
3966 len += sizeof(DATANAME);
3967 lpath = malloc(len);
3970 sprintf(lpath, "%s" DATANAME, path);
3973 /* The destination path must exist, but the destination file must not.
3974 * We don't want the OS to cache the writes, since the source data is
3975 * already in the OS cache.
3978 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
3979 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
3981 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL
3987 if (newfd == INVALID_HANDLE_VALUE) {
3992 #ifdef F_NOCACHE /* __APPLE__ */
3993 rc = fcntl(newfd, F_NOCACHE, 1);
4000 rc = mdb_env_copyfd(env, newfd);
4003 if (!(env->me_flags & MDB_NOSUBDIR))
4005 if (newfd != INVALID_HANDLE_VALUE)
4006 if (close(newfd) < 0 && rc == MDB_SUCCESS)
4013 mdb_env_close(MDB_env *env)
4020 VGMEMP_DESTROY(env);
4021 while ((dp = env->me_dpages) != NULL) {
4022 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
4023 env->me_dpages = dp->mp_next;
4027 mdb_env_close0(env, 0);
4031 /** Compare two items pointing at aligned size_t's */
4033 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
4035 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
4036 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
4039 /** Compare two items pointing at aligned int's */
4041 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
4043 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
4044 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
4047 /** Compare two items pointing at ints of unknown alignment.
4048 * Nodes and keys are guaranteed to be 2-byte aligned.
4051 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
4053 #if BYTE_ORDER == LITTLE_ENDIAN
4054 unsigned short *u, *c;
4057 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4058 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
4061 } while(!x && u > (unsigned short *)a->mv_data);
4064 return memcmp(a->mv_data, b->mv_data, a->mv_size);
4068 /** Compare two items lexically */
4070 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
4077 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4083 diff = memcmp(a->mv_data, b->mv_data, len);
4084 return diff ? diff : len_diff<0 ? -1 : len_diff;
4087 /** Compare two items in reverse byte order */
4089 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
4091 const unsigned char *p1, *p2, *p1_lim;
4095 p1_lim = (const unsigned char *)a->mv_data;
4096 p1 = (const unsigned char *)a->mv_data + a->mv_size;
4097 p2 = (const unsigned char *)b->mv_data + b->mv_size;
4099 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4105 while (p1 > p1_lim) {
4106 diff = *--p1 - *--p2;
4110 return len_diff<0 ? -1 : len_diff;
4113 /** Search for key within a page, using binary search.
4114 * Returns the smallest entry larger or equal to the key.
4115 * If exactp is non-null, stores whether the found entry was an exact match
4116 * in *exactp (1 or 0).
4117 * Updates the cursor index with the index of the found entry.
4118 * If no entry larger or equal to the key is found, returns NULL.
4121 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
4123 unsigned int i = 0, nkeys;
4126 MDB_page *mp = mc->mc_pg[mc->mc_top];
4127 MDB_node *node = NULL;
4132 nkeys = NUMKEYS(mp);
4137 COPY_PGNO(pgno, mp->mp_pgno);
4138 DPRINTF("searching %u keys in %s %spage %zu",
4139 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
4146 low = IS_LEAF(mp) ? 0 : 1;
4148 cmp = mc->mc_dbx->md_cmp;
4150 /* Branch pages have no data, so if using integer keys,
4151 * alignment is guaranteed. Use faster mdb_cmp_int.
4153 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
4154 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
4161 nodekey.mv_size = mc->mc_db->md_pad;
4162 node = NODEPTR(mp, 0); /* fake */
4163 while (low <= high) {
4164 i = (low + high) >> 1;
4165 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
4166 rc = cmp(key, &nodekey);
4167 DPRINTF("found leaf index %u [%s], rc = %i",
4168 i, DKEY(&nodekey), rc);
4177 while (low <= high) {
4178 i = (low + high) >> 1;
4180 node = NODEPTR(mp, i);
4181 nodekey.mv_size = NODEKSZ(node);
4182 nodekey.mv_data = NODEKEY(node);
4184 rc = cmp(key, &nodekey);
4187 DPRINTF("found leaf index %u [%s], rc = %i",
4188 i, DKEY(&nodekey), rc);
4190 DPRINTF("found branch index %u [%s -> %zu], rc = %i",
4191 i, DKEY(&nodekey), NODEPGNO(node), rc);
4202 if (rc > 0) { /* Found entry is less than the key. */
4203 i++; /* Skip to get the smallest entry larger than key. */
4205 node = NODEPTR(mp, i);
4208 *exactp = (rc == 0);
4209 /* store the key index */
4210 mc->mc_ki[mc->mc_top] = i;
4212 /* There is no entry larger or equal to the key. */
4215 /* nodeptr is fake for LEAF2 */
4221 mdb_cursor_adjust(MDB_cursor *mc, func)
4225 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
4226 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
4233 /** Pop a page off the top of the cursor's stack. */
4235 mdb_cursor_pop(MDB_cursor *mc)
4238 #ifndef MDB_DEBUG_SKIP
4239 MDB_page *top = mc->mc_pg[mc->mc_top];
4245 DPRINTF("popped page %zu off db %u cursor %p", top->mp_pgno,
4246 mc->mc_dbi, (void *) mc);
4250 /** Push a page onto the top of the cursor's stack. */
4252 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
4254 DPRINTF("pushing page %zu on db %u cursor %p", mp->mp_pgno,
4255 mc->mc_dbi, (void *) mc);
4257 if (mc->mc_snum >= CURSOR_STACK) {
4258 assert(mc->mc_snum < CURSOR_STACK);
4259 return MDB_CURSOR_FULL;
4262 mc->mc_top = mc->mc_snum++;
4263 mc->mc_pg[mc->mc_top] = mp;
4264 mc->mc_ki[mc->mc_top] = 0;
4269 /** Find the address of the page corresponding to a given page number.
4270 * @param[in] txn the transaction for this access.
4271 * @param[in] pgno the page number for the page to retrieve.
4272 * @param[out] ret address of a pointer where the page's address will be stored.
4273 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
4274 * @return 0 on success, non-zero on failure.
4277 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
4282 if (!((txn->mt_flags & MDB_TXN_RDONLY) |
4283 (txn->mt_env->me_flags & MDB_WRITEMAP)))
4288 MDB_ID2L dl = tx2->mt_u.dirty_list;
4290 /* Spilled pages were dirtied in this txn and flushed
4291 * because the dirty list got full. Bring this page
4292 * back in from the map (but don't unspill it here,
4293 * leave that unless page_touch happens again).
4295 if (tx2->mt_spill_pgs) {
4296 x = mdb_midl_search(tx2->mt_spill_pgs, pgno);
4297 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pgno) {
4298 p = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
4303 unsigned x = mdb_mid2l_search(dl, pgno);
4304 if (x <= dl[0].mid && dl[x].mid == pgno) {
4310 } while ((tx2 = tx2->mt_parent) != NULL);
4313 if (pgno < txn->mt_next_pgno) {
4315 p = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
4317 DPRINTF("page %zu not found", pgno);
4319 return MDB_PAGE_NOTFOUND;
4329 /** Search for the page a given key should be in.
4330 * Pushes parent pages on the cursor stack. This function continues a
4331 * search on a cursor that has already been initialized. (Usually by
4332 * #mdb_page_search() but also by #mdb_node_move().)
4333 * @param[in,out] mc the cursor for this operation.
4334 * @param[in] key the key to search for. If NULL, search for the lowest
4335 * page. (This is used by #mdb_cursor_first().)
4336 * @param[in] modify If true, visited pages are updated with new page numbers.
4337 * @return 0 on success, non-zero on failure.
4340 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int modify)
4342 MDB_page *mp = mc->mc_pg[mc->mc_top];
4347 while (IS_BRANCH(mp)) {
4351 DPRINTF("branch page %zu has %u keys", mp->mp_pgno, NUMKEYS(mp));
4352 assert(NUMKEYS(mp) > 1);
4353 DPRINTF("found index 0 to page %zu", NODEPGNO(NODEPTR(mp, 0)));
4355 if (key == NULL) /* Initialize cursor to first page. */
4357 else if (key->mv_size > MDB_MAXKEYSIZE && key->mv_data == NULL) {
4358 /* cursor to last page */
4362 node = mdb_node_search(mc, key, &exact);
4364 i = NUMKEYS(mp) - 1;
4366 i = mc->mc_ki[mc->mc_top];
4375 DPRINTF("following index %u for key [%s]",
4377 assert(i < NUMKEYS(mp));
4378 node = NODEPTR(mp, i);
4380 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4383 mc->mc_ki[mc->mc_top] = i;
4384 if ((rc = mdb_cursor_push(mc, mp)))
4388 if ((rc = mdb_page_touch(mc)) != 0)
4390 mp = mc->mc_pg[mc->mc_top];
4395 DPRINTF("internal error, index points to a %02X page!?",
4397 return MDB_CORRUPTED;
4400 DPRINTF("found leaf page %zu for key [%s]", mp->mp_pgno,
4401 key ? DKEY(key) : NULL);
4402 mc->mc_flags |= C_INITIALIZED;
4403 mc->mc_flags &= ~C_EOF;
4408 /** Search for the lowest key under the current branch page.
4409 * This just bypasses a NUMKEYS check in the current page
4410 * before calling mdb_page_search_root(), because the callers
4411 * are all in situations where the current page is known to
4415 mdb_page_search_lowest(MDB_cursor *mc)
4417 MDB_page *mp = mc->mc_pg[mc->mc_top];
4418 MDB_node *node = NODEPTR(mp, 0);
4421 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4424 mc->mc_ki[mc->mc_top] = 0;
4425 if ((rc = mdb_cursor_push(mc, mp)))
4427 return mdb_page_search_root(mc, NULL, 0);
4430 /** Search for the page a given key should be in.
4431 * Pushes parent pages on the cursor stack. This function just sets up
4432 * the search; it finds the root page for \b mc's database and sets this
4433 * as the root of the cursor's stack. Then #mdb_page_search_root() is
4434 * called to complete the search.
4435 * @param[in,out] mc the cursor for this operation.
4436 * @param[in] key the key to search for. If NULL, search for the lowest
4437 * page. (This is used by #mdb_cursor_first().)
4438 * @param[in] flags If MDB_PS_MODIFY set, visited pages are updated with new page numbers.
4439 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
4440 * @return 0 on success, non-zero on failure.
4443 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
4448 /* Make sure the txn is still viable, then find the root from
4449 * the txn's db table.
4451 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
4452 DPUTS("transaction has failed, must abort");
4455 /* Make sure we're using an up-to-date root */
4456 if (mc->mc_dbi > MAIN_DBI) {
4457 if ((*mc->mc_dbflag & DB_STALE) ||
4458 ((flags & MDB_PS_MODIFY) && !(*mc->mc_dbflag & DB_DIRTY))) {
4460 unsigned char dbflag = 0;
4461 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
4462 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, flags & MDB_PS_MODIFY);
4465 if (*mc->mc_dbflag & DB_STALE) {
4469 MDB_node *leaf = mdb_node_search(&mc2,
4470 &mc->mc_dbx->md_name, &exact);
4472 return MDB_NOTFOUND;
4473 rc = mdb_node_read(mc->mc_txn, leaf, &data);
4476 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
4478 /* The txn may not know this DBI, or another process may
4479 * have dropped and recreated the DB with other flags.
4481 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
4482 return MDB_INCOMPATIBLE;
4483 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
4485 if (flags & MDB_PS_MODIFY)
4487 *mc->mc_dbflag &= ~DB_STALE;
4488 *mc->mc_dbflag |= dbflag;
4491 root = mc->mc_db->md_root;
4493 if (root == P_INVALID) { /* Tree is empty. */
4494 DPUTS("tree is empty");
4495 return MDB_NOTFOUND;
4500 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
4501 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
4507 DPRINTF("db %u root page %zu has flags 0x%X",
4508 mc->mc_dbi, root, mc->mc_pg[0]->mp_flags);
4510 if (flags & MDB_PS_MODIFY) {
4511 if ((rc = mdb_page_touch(mc)))
4515 if (flags & MDB_PS_ROOTONLY)
4518 return mdb_page_search_root(mc, key, flags);
4522 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
4524 MDB_txn *txn = mc->mc_txn;
4525 pgno_t pg = mp->mp_pgno;
4526 unsigned i, ovpages = mp->mp_pages;
4527 MDB_env *env = txn->mt_env;
4530 DPRINTF("free ov page %zu (%d)", pg, ovpages);
4531 /* If the page is dirty or on the spill list we just acquired it,
4532 * so we should give it back to our current free list, if any.
4533 * Not currently supported in nested txns.
4534 * Otherwise put it onto the list of pages we freed in this txn.
4536 if (!(mp->mp_flags & P_DIRTY) && txn->mt_spill_pgs) {
4537 unsigned x = mdb_midl_search(txn->mt_spill_pgs, pg);
4538 if (x <= txn->mt_spill_pgs[0] && txn->mt_spill_pgs[x] == pg) {
4539 /* This page is no longer spilled */
4540 for (; x < txn->mt_spill_pgs[0]; x++)
4541 txn->mt_spill_pgs[x] = txn->mt_spill_pgs[x+1];
4542 txn->mt_spill_pgs[0]--;
4546 if ((mp->mp_flags & P_DIRTY) && !txn->mt_parent && env->me_pghead) {
4549 MDB_ID2 *dl, ix, iy;
4550 rc = mdb_midl_need(&env->me_pghead, ovpages);
4553 /* Remove from dirty list */
4554 dl = txn->mt_u.dirty_list;
4556 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
4564 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
4565 txn->mt_flags |= MDB_TXN_ERROR;
4566 return MDB_CORRUPTED;
4569 if (!(env->me_flags & MDB_WRITEMAP))
4570 mdb_dpage_free(env, mp);
4572 /* Insert in me_pghead */
4573 mop = env->me_pghead;
4574 j = mop[0] + ovpages;
4575 for (i = mop[0]; i && mop[i] < pg; i--)
4581 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
4585 mc->mc_db->md_overflow_pages -= ovpages;
4589 /** Return the data associated with a given node.
4590 * @param[in] txn The transaction for this operation.
4591 * @param[in] leaf The node being read.
4592 * @param[out] data Updated to point to the node's data.
4593 * @return 0 on success, non-zero on failure.
4596 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
4598 MDB_page *omp; /* overflow page */
4602 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
4603 data->mv_size = NODEDSZ(leaf);
4604 data->mv_data = NODEDATA(leaf);
4608 /* Read overflow data.
4610 data->mv_size = NODEDSZ(leaf);
4611 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
4612 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
4613 DPRINTF("read overflow page %zu failed", pgno);
4616 data->mv_data = METADATA(omp);
4622 mdb_get(MDB_txn *txn, MDB_dbi dbi,
4623 MDB_val *key, MDB_val *data)
4632 DPRINTF("===> get db %u key [%s]", dbi, DKEY(key));
4634 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
4637 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
4641 mdb_cursor_init(&mc, txn, dbi, &mx);
4642 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
4645 /** Find a sibling for a page.
4646 * Replaces the page at the top of the cursor's stack with the
4647 * specified sibling, if one exists.
4648 * @param[in] mc The cursor for this operation.
4649 * @param[in] move_right Non-zero if the right sibling is requested,
4650 * otherwise the left sibling.
4651 * @return 0 on success, non-zero on failure.
4654 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
4660 if (mc->mc_snum < 2) {
4661 return MDB_NOTFOUND; /* root has no siblings */
4665 DPRINTF("parent page is page %zu, index %u",
4666 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]);
4668 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
4669 : (mc->mc_ki[mc->mc_top] == 0)) {
4670 DPRINTF("no more keys left, moving to %s sibling",
4671 move_right ? "right" : "left");
4672 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
4673 /* undo cursor_pop before returning */
4680 mc->mc_ki[mc->mc_top]++;
4682 mc->mc_ki[mc->mc_top]--;
4683 DPRINTF("just moving to %s index key %u",
4684 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]);
4686 assert(IS_BRANCH(mc->mc_pg[mc->mc_top]));
4688 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4689 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL) != 0))
4692 mdb_cursor_push(mc, mp);
4694 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
4699 /** Move the cursor to the next data item. */
4701 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
4707 if (mc->mc_flags & C_EOF) {
4708 return MDB_NOTFOUND;
4711 assert(mc->mc_flags & C_INITIALIZED);
4713 mp = mc->mc_pg[mc->mc_top];
4715 if (mc->mc_db->md_flags & MDB_DUPSORT) {
4716 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4717 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4718 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
4719 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
4720 if (op != MDB_NEXT || rc != MDB_NOTFOUND)
4724 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
4725 if (op == MDB_NEXT_DUP)
4726 return MDB_NOTFOUND;
4730 DPRINTF("cursor_next: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
4732 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
4733 DPUTS("=====> move to next sibling page");
4734 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
4735 mc->mc_flags |= C_EOF;
4738 mp = mc->mc_pg[mc->mc_top];
4739 DPRINTF("next page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
4741 mc->mc_ki[mc->mc_top]++;
4743 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
4744 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
4747 key->mv_size = mc->mc_db->md_pad;
4748 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4752 assert(IS_LEAF(mp));
4753 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4755 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4756 mdb_xcursor_init1(mc, leaf);
4759 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4762 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4763 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4764 if (rc != MDB_SUCCESS)
4769 MDB_GET_KEY(leaf, key);
4773 /** Move the cursor to the previous data item. */
4775 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
4781 assert(mc->mc_flags & C_INITIALIZED);
4783 mp = mc->mc_pg[mc->mc_top];
4785 if (mc->mc_db->md_flags & MDB_DUPSORT) {
4786 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4787 if (op == MDB_PREV || op == MDB_PREV_DUP) {
4788 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4789 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
4790 if (op != MDB_PREV || rc != MDB_NOTFOUND)
4793 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
4794 if (op == MDB_PREV_DUP)
4795 return MDB_NOTFOUND;
4800 DPRINTF("cursor_prev: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
4802 if (mc->mc_ki[mc->mc_top] == 0) {
4803 DPUTS("=====> move to prev sibling page");
4804 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
4807 mp = mc->mc_pg[mc->mc_top];
4808 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
4809 DPRINTF("prev page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
4811 mc->mc_ki[mc->mc_top]--;
4813 mc->mc_flags &= ~C_EOF;
4815 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
4816 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
4819 key->mv_size = mc->mc_db->md_pad;
4820 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4824 assert(IS_LEAF(mp));
4825 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4827 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4828 mdb_xcursor_init1(mc, leaf);
4831 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4834 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4835 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4836 if (rc != MDB_SUCCESS)
4841 MDB_GET_KEY(leaf, key);
4845 /** Set the cursor on a specific data item. */
4847 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4848 MDB_cursor_op op, int *exactp)
4852 MDB_node *leaf = NULL;
4857 assert(key->mv_size > 0);
4859 if (mc->mc_db->md_flags & MDB_DUPSORT)
4860 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
4862 /* See if we're already on the right page */
4863 if (mc->mc_flags & C_INITIALIZED) {
4866 mp = mc->mc_pg[mc->mc_top];
4868 mc->mc_ki[mc->mc_top] = 0;
4869 return MDB_NOTFOUND;
4871 if (mp->mp_flags & P_LEAF2) {
4872 nodekey.mv_size = mc->mc_db->md_pad;
4873 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
4875 leaf = NODEPTR(mp, 0);
4876 MDB_GET_KEY(leaf, &nodekey);
4878 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4880 /* Probably happens rarely, but first node on the page
4881 * was the one we wanted.
4883 mc->mc_ki[mc->mc_top] = 0;
4890 unsigned int nkeys = NUMKEYS(mp);
4892 if (mp->mp_flags & P_LEAF2) {
4893 nodekey.mv_data = LEAF2KEY(mp,
4894 nkeys-1, nodekey.mv_size);
4896 leaf = NODEPTR(mp, nkeys-1);
4897 MDB_GET_KEY(leaf, &nodekey);
4899 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4901 /* last node was the one we wanted */
4902 mc->mc_ki[mc->mc_top] = nkeys-1;
4908 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
4909 /* This is definitely the right page, skip search_page */
4910 if (mp->mp_flags & P_LEAF2) {
4911 nodekey.mv_data = LEAF2KEY(mp,
4912 mc->mc_ki[mc->mc_top], nodekey.mv_size);
4914 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4915 MDB_GET_KEY(leaf, &nodekey);
4917 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4919 /* current node was the one we wanted */
4929 /* If any parents have right-sibs, search.
4930 * Otherwise, there's nothing further.
4932 for (i=0; i<mc->mc_top; i++)
4934 NUMKEYS(mc->mc_pg[i])-1)
4936 if (i == mc->mc_top) {
4937 /* There are no other pages */
4938 mc->mc_ki[mc->mc_top] = nkeys;
4939 return MDB_NOTFOUND;
4943 /* There are no other pages */
4944 mc->mc_ki[mc->mc_top] = 0;
4945 return MDB_NOTFOUND;
4949 rc = mdb_page_search(mc, key, 0);
4950 if (rc != MDB_SUCCESS)
4953 mp = mc->mc_pg[mc->mc_top];
4954 assert(IS_LEAF(mp));
4957 leaf = mdb_node_search(mc, key, exactp);
4958 if (exactp != NULL && !*exactp) {
4959 /* MDB_SET specified and not an exact match. */
4960 return MDB_NOTFOUND;
4964 DPUTS("===> inexact leaf not found, goto sibling");
4965 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
4966 return rc; /* no entries matched */
4967 mp = mc->mc_pg[mc->mc_top];
4968 assert(IS_LEAF(mp));
4969 leaf = NODEPTR(mp, 0);
4973 mc->mc_flags |= C_INITIALIZED;
4974 mc->mc_flags &= ~C_EOF;
4977 key->mv_size = mc->mc_db->md_pad;
4978 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4982 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4983 mdb_xcursor_init1(mc, leaf);
4986 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4987 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
4988 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4991 if (op == MDB_GET_BOTH) {
4997 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
4998 if (rc != MDB_SUCCESS)
5001 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
5003 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
5005 rc = mc->mc_dbx->md_dcmp(data, &d2);
5007 if (op == MDB_GET_BOTH || rc > 0)
5008 return MDB_NOTFOUND;
5013 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5014 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5019 /* The key already matches in all other cases */
5020 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
5021 MDB_GET_KEY(leaf, key);
5022 DPRINTF("==> cursor placed on key [%s]", DKEY(key));
5027 /** Move the cursor to the first item in the database. */
5029 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5034 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5035 rc = mdb_page_search(mc, NULL, 0);
5036 if (rc != MDB_SUCCESS)
5039 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
5041 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
5042 mc->mc_flags |= C_INITIALIZED;
5043 mc->mc_flags &= ~C_EOF;
5045 mc->mc_ki[mc->mc_top] = 0;
5047 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5048 key->mv_size = mc->mc_db->md_pad;
5049 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
5054 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5055 mdb_xcursor_init1(mc, leaf);
5056 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5061 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5062 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5066 MDB_GET_KEY(leaf, key);
5070 /** Move the cursor to the last item in the database. */
5072 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5077 if (!(mc->mc_flags & C_EOF)) {
5079 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5082 lkey.mv_size = MDB_MAXKEYSIZE+1;
5083 lkey.mv_data = NULL;
5084 rc = mdb_page_search(mc, &lkey, 0);
5085 if (rc != MDB_SUCCESS)
5088 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
5091 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
5092 mc->mc_flags |= C_INITIALIZED|C_EOF;
5093 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5095 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5096 key->mv_size = mc->mc_db->md_pad;
5097 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
5102 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5103 mdb_xcursor_init1(mc, leaf);
5104 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5109 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5110 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5115 MDB_GET_KEY(leaf, key);
5120 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5129 case MDB_GET_CURRENT:
5130 if (!(mc->mc_flags & C_INITIALIZED)) {
5133 MDB_page *mp = mc->mc_pg[mc->mc_top];
5135 mc->mc_ki[mc->mc_top] = 0;
5141 key->mv_size = mc->mc_db->md_pad;
5142 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5144 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5145 MDB_GET_KEY(leaf, key);
5147 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5148 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
5150 rc = mdb_node_read(mc->mc_txn, leaf, data);
5157 case MDB_GET_BOTH_RANGE:
5158 if (data == NULL || mc->mc_xcursor == NULL) {
5166 if (key == NULL || key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
5168 } else if (op == MDB_SET_RANGE)
5169 rc = mdb_cursor_set(mc, key, data, op, NULL);
5171 rc = mdb_cursor_set(mc, key, data, op, &exact);
5173 case MDB_GET_MULTIPLE:
5175 !(mc->mc_db->md_flags & MDB_DUPFIXED) ||
5176 !(mc->mc_flags & C_INITIALIZED)) {
5181 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
5182 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
5185 case MDB_NEXT_MULTIPLE:
5187 !(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5191 if (!(mc->mc_flags & C_INITIALIZED))
5192 rc = mdb_cursor_first(mc, key, data);
5194 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
5195 if (rc == MDB_SUCCESS) {
5196 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
5199 mx = &mc->mc_xcursor->mx_cursor;
5200 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
5202 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
5203 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
5211 case MDB_NEXT_NODUP:
5212 if (!(mc->mc_flags & C_INITIALIZED))
5213 rc = mdb_cursor_first(mc, key, data);
5215 rc = mdb_cursor_next(mc, key, data, op);
5219 case MDB_PREV_NODUP:
5220 if (!(mc->mc_flags & C_INITIALIZED)) {
5221 rc = mdb_cursor_last(mc, key, data);
5224 mc->mc_flags |= C_INITIALIZED;
5225 mc->mc_ki[mc->mc_top]++;
5227 rc = mdb_cursor_prev(mc, key, data, op);
5230 rc = mdb_cursor_first(mc, key, data);
5234 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
5235 !(mc->mc_flags & C_INITIALIZED) ||
5236 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
5240 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5243 rc = mdb_cursor_last(mc, key, data);
5247 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
5248 !(mc->mc_flags & C_INITIALIZED) ||
5249 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
5253 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5256 DPRINTF("unhandled/unimplemented cursor operation %u", op);
5264 /** Touch all the pages in the cursor stack.
5265 * Makes sure all the pages are writable, before attempting a write operation.
5266 * @param[in] mc The cursor to operate on.
5269 mdb_cursor_touch(MDB_cursor *mc)
5273 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
5276 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
5277 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
5280 *mc->mc_dbflag |= DB_DIRTY;
5282 for (mc->mc_top = 0; mc->mc_top < mc->mc_snum; mc->mc_top++) {
5283 rc = mdb_page_touch(mc);
5287 mc->mc_top = mc->mc_snum-1;
5291 /** Do not spill pages to disk if txn is getting full, may fail instead */
5292 #define MDB_NOSPILL 0x8000
5295 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5298 enum { MDB_NO_ROOT = MDB_LAST_ERRCODE+10 }; /* internal code */
5299 MDB_node *leaf = NULL;
5300 MDB_val xdata, *rdata, dkey;
5303 int do_sub = 0, insert = 0;
5304 unsigned int mcount = 0, dcount = 0, nospill;
5308 char dbuf[MDB_MAXKEYSIZE+1];
5309 unsigned int nflags;
5312 /* Check this first so counter will always be zero on any
5315 if (flags & MDB_MULTIPLE) {
5316 dcount = data[1].mv_size;
5317 data[1].mv_size = 0;
5318 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
5322 nospill = flags & MDB_NOSPILL;
5323 flags &= ~MDB_NOSPILL;
5325 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
5328 if (flags != MDB_CURRENT && (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE))
5331 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT) && data->mv_size > MDB_MAXKEYSIZE)
5334 #if SIZE_MAX > MAXDATASIZE
5335 if (data->mv_size > MAXDATASIZE)
5339 DPRINTF("==> put db %u key [%s], size %zu, data size %zu",
5340 mc->mc_dbi, DKEY(key), key ? key->mv_size:0, data->mv_size);
5344 if (flags == MDB_CURRENT) {
5345 if (!(mc->mc_flags & C_INITIALIZED))
5348 } else if (mc->mc_db->md_root == P_INVALID) {
5349 /* new database, cursor has nothing to point to */
5351 mc->mc_flags &= ~C_INITIALIZED;
5356 if (flags & MDB_APPEND) {
5358 rc = mdb_cursor_last(mc, &k2, &d2);
5360 rc = mc->mc_dbx->md_cmp(key, &k2);
5363 mc->mc_ki[mc->mc_top]++;
5365 /* new key is <= last key */
5370 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
5372 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
5373 DPRINTF("duplicate key [%s]", DKEY(key));
5375 return MDB_KEYEXIST;
5377 if (rc && rc != MDB_NOTFOUND)
5381 /* Cursor is positioned, check for room in the dirty list */
5383 if (flags & MDB_MULTIPLE) {
5385 xdata.mv_size = data->mv_size * dcount;
5389 if ((rc2 = mdb_page_spill(mc, key, rdata)))
5393 if (rc == MDB_NO_ROOT) {
5395 /* new database, write a root leaf page */
5396 DPUTS("allocating new root leaf page");
5397 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) {
5400 mdb_cursor_push(mc, np);
5401 mc->mc_db->md_root = np->mp_pgno;
5402 mc->mc_db->md_depth++;
5403 *mc->mc_dbflag |= DB_DIRTY;
5404 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
5406 np->mp_flags |= P_LEAF2;
5407 mc->mc_flags |= C_INITIALIZED;
5409 /* make sure all cursor pages are writable */
5410 rc2 = mdb_cursor_touch(mc);
5415 /* The key already exists */
5416 if (rc == MDB_SUCCESS) {
5417 /* there's only a key anyway, so this is a no-op */
5418 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5419 unsigned int ksize = mc->mc_db->md_pad;
5420 if (key->mv_size != ksize)
5422 if (flags == MDB_CURRENT) {
5423 char *ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
5424 memcpy(ptr, key->mv_data, ksize);
5429 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5432 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
5433 /* Was a single item before, must convert now */
5435 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5436 /* Just overwrite the current item */
5437 if (flags == MDB_CURRENT)
5440 dkey.mv_size = NODEDSZ(leaf);
5441 dkey.mv_data = NODEDATA(leaf);
5442 #if UINT_MAX < SIZE_MAX
5443 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && dkey.mv_size == sizeof(size_t))
5444 #ifdef MISALIGNED_OK
5445 mc->mc_dbx->md_dcmp = mdb_cmp_long;
5447 mc->mc_dbx->md_dcmp = mdb_cmp_cint;
5450 /* if data matches, ignore it */
5451 if (!mc->mc_dbx->md_dcmp(data, &dkey))
5452 return (flags == MDB_NODUPDATA) ? MDB_KEYEXIST : MDB_SUCCESS;
5454 /* create a fake page for the dup items */
5455 memcpy(dbuf, dkey.mv_data, dkey.mv_size);
5456 dkey.mv_data = dbuf;
5457 fp = (MDB_page *)&pbuf;
5458 fp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5459 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
5460 fp->mp_lower = PAGEHDRSZ;
5461 fp->mp_upper = PAGEHDRSZ + dkey.mv_size + data->mv_size;
5462 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5463 fp->mp_flags |= P_LEAF2;
5464 fp->mp_pad = data->mv_size;
5465 fp->mp_upper += 2 * data->mv_size; /* leave space for 2 more */
5467 fp->mp_upper += 2 * sizeof(indx_t) + 2 * NODESIZE +
5468 (dkey.mv_size & 1) + (data->mv_size & 1);
5470 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5473 xdata.mv_size = fp->mp_upper;
5478 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
5479 /* See if we need to convert from fake page to subDB */
5481 unsigned int offset;
5485 fp = NODEDATA(leaf);
5486 if (flags == MDB_CURRENT) {
5488 fp->mp_flags |= P_DIRTY;
5489 COPY_PGNO(fp->mp_pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
5490 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
5494 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5495 offset = fp->mp_pad;
5496 if (SIZELEFT(fp) >= offset)
5498 offset *= 4; /* space for 4 more */
5500 offset = NODESIZE + sizeof(indx_t) + data->mv_size;
5502 offset += offset & 1;
5503 fp_flags = fp->mp_flags;
5504 if (NODESIZE + sizeof(indx_t) + NODEKSZ(leaf) + NODEDSZ(leaf) +
5505 offset >= mc->mc_txn->mt_env->me_nodemax) {
5506 /* yes, convert it */
5508 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5509 dummy.md_pad = fp->mp_pad;
5510 dummy.md_flags = MDB_DUPFIXED;
5511 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
5512 dummy.md_flags |= MDB_INTEGERKEY;
5515 dummy.md_branch_pages = 0;
5516 dummy.md_leaf_pages = 1;
5517 dummy.md_overflow_pages = 0;
5518 dummy.md_entries = NUMKEYS(fp);
5520 xdata.mv_size = sizeof(MDB_db);
5521 xdata.mv_data = &dummy;
5522 if ((rc = mdb_page_alloc(mc, 1, &mp)))
5524 offset = mc->mc_txn->mt_env->me_psize - NODEDSZ(leaf);
5525 flags |= F_DUPDATA|F_SUBDATA;
5526 dummy.md_root = mp->mp_pgno;
5527 fp_flags &= ~P_SUBP;
5529 /* no, just grow it */
5531 xdata.mv_size = NODEDSZ(leaf) + offset;
5532 xdata.mv_data = &pbuf;
5533 mp = (MDB_page *)&pbuf;
5534 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5537 mp->mp_flags = fp_flags | P_DIRTY;
5538 mp->mp_pad = fp->mp_pad;
5539 mp->mp_lower = fp->mp_lower;
5540 mp->mp_upper = fp->mp_upper + offset;
5542 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
5544 nsize = NODEDSZ(leaf) - fp->mp_upper;
5545 memcpy((char *)mp + mp->mp_upper, (char *)fp + fp->mp_upper, nsize);
5546 for (i=0; i<NUMKEYS(fp); i++)
5547 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
5549 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5553 /* data is on sub-DB, just store it */
5554 flags |= F_DUPDATA|F_SUBDATA;
5558 /* overflow page overwrites need special handling */
5559 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5562 unsigned psize = mc->mc_txn->mt_env->me_psize;
5563 int level, ovpages, dpages = OVPAGES(data->mv_size, psize);
5565 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
5566 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
5568 ovpages = omp->mp_pages;
5570 /* Is the ov page large enough? */
5571 if (ovpages >= dpages) {
5572 if (!(omp->mp_flags & P_DIRTY) &&
5573 (level || (mc->mc_txn->mt_env->me_flags & MDB_WRITEMAP)))
5575 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
5578 level = 0; /* dirty in this txn or clean */
5581 if (omp->mp_flags & P_DIRTY) {
5582 /* yes, overwrite it. Note in this case we don't
5583 * bother to try shrinking the page if the new data
5584 * is smaller than the overflow threshold.
5587 /* It is writable only in a parent txn */
5588 size_t sz = (size_t) psize * ovpages, off;
5589 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
5595 mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
5596 if (!(flags & MDB_RESERVE)) {
5597 /* Copy end of page, adjusting alignment so
5598 * compiler may copy words instead of bytes.
5600 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
5601 memcpy((size_t *)((char *)np + off),
5602 (size_t *)((char *)omp + off), sz - off);
5605 memcpy(np, omp, sz); /* Copy beginning of page */
5608 SETDSZ(leaf, data->mv_size);
5609 if (F_ISSET(flags, MDB_RESERVE))
5610 data->mv_data = METADATA(omp);
5612 memcpy(METADATA(omp), data->mv_data, data->mv_size);
5616 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
5618 } else if (NODEDSZ(leaf) == data->mv_size) {
5619 /* same size, just replace it. Note that we could
5620 * also reuse this node if the new data is smaller,
5621 * but instead we opt to shrink the node in that case.
5623 if (F_ISSET(flags, MDB_RESERVE))
5624 data->mv_data = NODEDATA(leaf);
5625 else if (data->mv_size)
5626 memcpy(NODEDATA(leaf), data->mv_data, data->mv_size);
5628 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
5631 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5632 mc->mc_db->md_entries--;
5634 DPRINTF("inserting key at index %i", mc->mc_ki[mc->mc_top]);
5641 nflags = flags & NODE_ADD_FLAGS;
5642 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(mc->mc_txn->mt_env, key, rdata);
5643 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
5644 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
5645 nflags &= ~MDB_APPEND;
5647 nflags |= MDB_SPLIT_REPLACE;
5648 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
5650 /* There is room already in this leaf page. */
5651 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
5652 if (rc == 0 && !do_sub && insert) {
5653 /* Adjust other cursors pointing to mp */
5654 MDB_cursor *m2, *m3;
5655 MDB_dbi dbi = mc->mc_dbi;
5656 unsigned i = mc->mc_top;
5657 MDB_page *mp = mc->mc_pg[i];
5659 if (mc->mc_flags & C_SUB)
5662 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5663 if (mc->mc_flags & C_SUB)
5664 m3 = &m2->mc_xcursor->mx_cursor;
5667 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
5668 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
5675 if (rc != MDB_SUCCESS)
5676 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5678 /* Now store the actual data in the child DB. Note that we're
5679 * storing the user data in the keys field, so there are strict
5680 * size limits on dupdata. The actual data fields of the child
5681 * DB are all zero size.
5688 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5689 if (flags & MDB_CURRENT) {
5690 xflags = MDB_CURRENT|MDB_NOSPILL;
5692 mdb_xcursor_init1(mc, leaf);
5693 xflags = (flags & MDB_NODUPDATA) ?
5694 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
5696 /* converted, write the original data first */
5698 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
5702 /* Adjust other cursors pointing to mp */
5704 unsigned i = mc->mc_top;
5705 MDB_page *mp = mc->mc_pg[i];
5707 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5708 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
5709 if (!(m2->mc_flags & C_INITIALIZED)) continue;
5710 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
5711 mdb_xcursor_init1(m2, leaf);
5715 /* we've done our job */
5718 if (flags & MDB_APPENDDUP)
5719 xflags |= MDB_APPEND;
5720 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
5721 if (flags & F_SUBDATA) {
5722 void *db = NODEDATA(leaf);
5723 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
5726 /* sub-writes might have failed so check rc again.
5727 * Don't increment count if we just replaced an existing item.
5729 if (!rc && !(flags & MDB_CURRENT))
5730 mc->mc_db->md_entries++;
5731 if (flags & MDB_MULTIPLE) {
5734 if (mcount < dcount) {
5735 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
5736 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5740 /* let caller know how many succeeded, if any */
5741 data[1].mv_size = mcount;
5745 /* If we succeeded and the key didn't exist before, make sure
5746 * the cursor is marked valid.
5749 mc->mc_flags |= C_INITIALIZED;
5754 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
5759 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
5762 if (!(mc->mc_flags & C_INITIALIZED))
5765 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
5767 flags &= ~MDB_NOSPILL; /* TODO: Or change (flags != MDB_NODUPDATA) to ~(flags & MDB_NODUPDATA), not looking at the logic of that code just now */
5769 rc = mdb_cursor_touch(mc);
5773 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5775 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5776 if (flags != MDB_NODUPDATA) {
5777 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
5778 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
5780 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
5781 /* If sub-DB still has entries, we're done */
5782 if (mc->mc_xcursor->mx_db.md_entries) {
5783 if (leaf->mn_flags & F_SUBDATA) {
5784 /* update subDB info */
5785 void *db = NODEDATA(leaf);
5786 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
5789 /* shrink fake page */
5790 mdb_node_shrink(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5791 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5792 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
5793 /* fix other sub-DB cursors pointed at this fake page */
5794 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5795 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
5796 if (m2->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top] &&
5797 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
5798 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
5801 mc->mc_db->md_entries--;
5804 /* otherwise fall thru and delete the sub-DB */
5807 if (leaf->mn_flags & F_SUBDATA) {
5808 /* add all the child DB's pages to the free list */
5809 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
5810 if (rc == MDB_SUCCESS) {
5811 mc->mc_db->md_entries -=
5812 mc->mc_xcursor->mx_db.md_entries;
5817 return mdb_cursor_del0(mc, leaf);
5820 /** Allocate and initialize new pages for a database.
5821 * @param[in] mc a cursor on the database being added to.
5822 * @param[in] flags flags defining what type of page is being allocated.
5823 * @param[in] num the number of pages to allocate. This is usually 1,
5824 * unless allocating overflow pages for a large record.
5825 * @param[out] mp Address of a page, or NULL on failure.
5826 * @return 0 on success, non-zero on failure.
5829 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
5834 if ((rc = mdb_page_alloc(mc, num, &np)))
5836 DPRINTF("allocated new mpage %zu, page size %u",
5837 np->mp_pgno, mc->mc_txn->mt_env->me_psize);
5838 np->mp_flags = flags | P_DIRTY;
5839 np->mp_lower = PAGEHDRSZ;
5840 np->mp_upper = mc->mc_txn->mt_env->me_psize;
5843 mc->mc_db->md_branch_pages++;
5844 else if (IS_LEAF(np))
5845 mc->mc_db->md_leaf_pages++;
5846 else if (IS_OVERFLOW(np)) {
5847 mc->mc_db->md_overflow_pages += num;
5855 /** Calculate the size of a leaf node.
5856 * The size depends on the environment's page size; if a data item
5857 * is too large it will be put onto an overflow page and the node
5858 * size will only include the key and not the data. Sizes are always
5859 * rounded up to an even number of bytes, to guarantee 2-byte alignment
5860 * of the #MDB_node headers.
5861 * @param[in] env The environment handle.
5862 * @param[in] key The key for the node.
5863 * @param[in] data The data for the node.
5864 * @return The number of bytes needed to store the node.
5867 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
5871 sz = LEAFSIZE(key, data);
5872 if (sz >= env->me_nodemax) {
5873 /* put on overflow page */
5874 sz -= data->mv_size - sizeof(pgno_t);
5878 return sz + sizeof(indx_t);
5881 /** Calculate the size of a branch node.
5882 * The size should depend on the environment's page size but since
5883 * we currently don't support spilling large keys onto overflow
5884 * pages, it's simply the size of the #MDB_node header plus the
5885 * size of the key. Sizes are always rounded up to an even number
5886 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
5887 * @param[in] env The environment handle.
5888 * @param[in] key The key for the node.
5889 * @return The number of bytes needed to store the node.
5892 mdb_branch_size(MDB_env *env, MDB_val *key)
5897 if (sz >= env->me_nodemax) {
5898 /* put on overflow page */
5899 /* not implemented */
5900 /* sz -= key->size - sizeof(pgno_t); */
5903 return sz + sizeof(indx_t);
5906 /** Add a node to the page pointed to by the cursor.
5907 * @param[in] mc The cursor for this operation.
5908 * @param[in] indx The index on the page where the new node should be added.
5909 * @param[in] key The key for the new node.
5910 * @param[in] data The data for the new node, if any.
5911 * @param[in] pgno The page number, if adding a branch node.
5912 * @param[in] flags Flags for the node.
5913 * @return 0 on success, non-zero on failure. Possible errors are:
5915 * <li>ENOMEM - failed to allocate overflow pages for the node.
5916 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
5917 * should never happen since all callers already calculate the
5918 * page's free space before calling this function.
5922 mdb_node_add(MDB_cursor *mc, indx_t indx,
5923 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
5926 size_t node_size = NODESIZE;
5929 MDB_page *mp = mc->mc_pg[mc->mc_top];
5930 MDB_page *ofp = NULL; /* overflow page */
5933 assert(mp->mp_upper >= mp->mp_lower);
5935 DPRINTF("add to %s %spage %zu index %i, data size %zu key size %zu [%s]",
5936 IS_LEAF(mp) ? "leaf" : "branch",
5937 IS_SUBP(mp) ? "sub-" : "",
5938 mp->mp_pgno, indx, data ? data->mv_size : 0,
5939 key ? key->mv_size : 0, key ? DKEY(key) : NULL);
5942 /* Move higher keys up one slot. */
5943 int ksize = mc->mc_db->md_pad, dif;
5944 char *ptr = LEAF2KEY(mp, indx, ksize);
5945 dif = NUMKEYS(mp) - indx;
5947 memmove(ptr+ksize, ptr, dif*ksize);
5948 /* insert new key */
5949 memcpy(ptr, key->mv_data, ksize);
5951 /* Just using these for counting */
5952 mp->mp_lower += sizeof(indx_t);
5953 mp->mp_upper -= ksize - sizeof(indx_t);
5958 node_size += key->mv_size;
5962 if (F_ISSET(flags, F_BIGDATA)) {
5963 /* Data already on overflow page. */
5964 node_size += sizeof(pgno_t);
5965 } else if (node_size + data->mv_size >= mc->mc_txn->mt_env->me_nodemax) {
5966 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
5968 /* Put data on overflow page. */
5969 DPRINTF("data size is %zu, node would be %zu, put data on overflow page",
5970 data->mv_size, node_size+data->mv_size);
5971 node_size += sizeof(pgno_t);
5972 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
5974 DPRINTF("allocated overflow page %zu", ofp->mp_pgno);
5977 node_size += data->mv_size;
5980 node_size += node_size & 1;
5982 if (node_size + sizeof(indx_t) > SIZELEFT(mp)) {
5983 DPRINTF("not enough room in page %zu, got %u ptrs",
5984 mp->mp_pgno, NUMKEYS(mp));
5985 DPRINTF("upper - lower = %u - %u = %u", mp->mp_upper, mp->mp_lower,
5986 mp->mp_upper - mp->mp_lower);
5987 DPRINTF("node size = %zu", node_size);
5988 return MDB_PAGE_FULL;
5991 /* Move higher pointers up one slot. */
5992 for (i = NUMKEYS(mp); i > indx; i--)
5993 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
5995 /* Adjust free space offsets. */
5996 ofs = mp->mp_upper - node_size;
5997 assert(ofs >= mp->mp_lower + sizeof(indx_t));
5998 mp->mp_ptrs[indx] = ofs;
6000 mp->mp_lower += sizeof(indx_t);
6002 /* Write the node data. */
6003 node = NODEPTR(mp, indx);
6004 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
6005 node->mn_flags = flags;
6007 SETDSZ(node,data->mv_size);
6012 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6017 if (F_ISSET(flags, F_BIGDATA))
6018 memcpy(node->mn_data + key->mv_size, data->mv_data,
6020 else if (F_ISSET(flags, MDB_RESERVE))
6021 data->mv_data = node->mn_data + key->mv_size;
6023 memcpy(node->mn_data + key->mv_size, data->mv_data,
6026 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
6028 if (F_ISSET(flags, MDB_RESERVE))
6029 data->mv_data = METADATA(ofp);
6031 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
6038 /** Delete the specified node from a page.
6039 * @param[in] mp The page to operate on.
6040 * @param[in] indx The index of the node to delete.
6041 * @param[in] ksize The size of a node. Only used if the page is
6042 * part of a #MDB_DUPFIXED database.
6045 mdb_node_del(MDB_page *mp, indx_t indx, int ksize)
6048 indx_t i, j, numkeys, ptr;
6055 COPY_PGNO(pgno, mp->mp_pgno);
6056 DPRINTF("delete node %u on %s page %zu", indx,
6057 IS_LEAF(mp) ? "leaf" : "branch", pgno);
6060 assert(indx < NUMKEYS(mp));
6063 int x = NUMKEYS(mp) - 1 - indx;
6064 base = LEAF2KEY(mp, indx, ksize);
6066 memmove(base, base + ksize, x * ksize);
6067 mp->mp_lower -= sizeof(indx_t);
6068 mp->mp_upper += ksize - sizeof(indx_t);
6072 node = NODEPTR(mp, indx);
6073 sz = NODESIZE + node->mn_ksize;
6075 if (F_ISSET(node->mn_flags, F_BIGDATA))
6076 sz += sizeof(pgno_t);
6078 sz += NODEDSZ(node);
6082 ptr = mp->mp_ptrs[indx];
6083 numkeys = NUMKEYS(mp);
6084 for (i = j = 0; i < numkeys; i++) {
6086 mp->mp_ptrs[j] = mp->mp_ptrs[i];
6087 if (mp->mp_ptrs[i] < ptr)
6088 mp->mp_ptrs[j] += sz;
6093 base = (char *)mp + mp->mp_upper;
6094 memmove(base + sz, base, ptr - mp->mp_upper);
6096 mp->mp_lower -= sizeof(indx_t);
6100 /** Compact the main page after deleting a node on a subpage.
6101 * @param[in] mp The main page to operate on.
6102 * @param[in] indx The index of the subpage on the main page.
6105 mdb_node_shrink(MDB_page *mp, indx_t indx)
6112 indx_t i, numkeys, ptr;
6114 node = NODEPTR(mp, indx);
6115 sp = (MDB_page *)NODEDATA(node);
6116 osize = NODEDSZ(node);
6118 delta = sp->mp_upper - sp->mp_lower;
6119 SETDSZ(node, osize - delta);
6120 xp = (MDB_page *)((char *)sp + delta);
6122 /* shift subpage upward */
6124 nsize = NUMKEYS(sp) * sp->mp_pad;
6125 memmove(METADATA(xp), METADATA(sp), nsize);
6128 nsize = osize - sp->mp_upper;
6129 numkeys = NUMKEYS(sp);
6130 for (i=numkeys-1; i>=0; i--)
6131 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
6133 xp->mp_upper = sp->mp_lower;
6134 xp->mp_lower = sp->mp_lower;
6135 xp->mp_flags = sp->mp_flags;
6136 xp->mp_pad = sp->mp_pad;
6137 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
6139 /* shift lower nodes upward */
6140 ptr = mp->mp_ptrs[indx];
6141 numkeys = NUMKEYS(mp);
6142 for (i = 0; i < numkeys; i++) {
6143 if (mp->mp_ptrs[i] <= ptr)
6144 mp->mp_ptrs[i] += delta;
6147 base = (char *)mp + mp->mp_upper;
6148 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
6149 mp->mp_upper += delta;
6152 /** Initial setup of a sorted-dups cursor.
6153 * Sorted duplicates are implemented as a sub-database for the given key.
6154 * The duplicate data items are actually keys of the sub-database.
6155 * Operations on the duplicate data items are performed using a sub-cursor
6156 * initialized when the sub-database is first accessed. This function does
6157 * the preliminary setup of the sub-cursor, filling in the fields that
6158 * depend only on the parent DB.
6159 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6162 mdb_xcursor_init0(MDB_cursor *mc)
6164 MDB_xcursor *mx = mc->mc_xcursor;
6166 mx->mx_cursor.mc_xcursor = NULL;
6167 mx->mx_cursor.mc_txn = mc->mc_txn;
6168 mx->mx_cursor.mc_db = &mx->mx_db;
6169 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
6170 mx->mx_cursor.mc_dbi = mc->mc_dbi+1;
6171 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
6172 mx->mx_cursor.mc_snum = 0;
6173 mx->mx_cursor.mc_top = 0;
6174 mx->mx_cursor.mc_flags = C_SUB;
6175 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
6176 mx->mx_dbx.md_dcmp = NULL;
6177 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
6180 /** Final setup of a sorted-dups cursor.
6181 * Sets up the fields that depend on the data from the main cursor.
6182 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6183 * @param[in] node The data containing the #MDB_db record for the
6184 * sorted-dup database.
6187 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
6189 MDB_xcursor *mx = mc->mc_xcursor;
6191 if (node->mn_flags & F_SUBDATA) {
6192 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
6193 mx->mx_cursor.mc_pg[0] = 0;
6194 mx->mx_cursor.mc_snum = 0;
6195 mx->mx_cursor.mc_flags = C_SUB;
6197 MDB_page *fp = NODEDATA(node);
6198 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
6199 mx->mx_db.md_flags = 0;
6200 mx->mx_db.md_depth = 1;
6201 mx->mx_db.md_branch_pages = 0;
6202 mx->mx_db.md_leaf_pages = 1;
6203 mx->mx_db.md_overflow_pages = 0;
6204 mx->mx_db.md_entries = NUMKEYS(fp);
6205 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
6206 mx->mx_cursor.mc_snum = 1;
6207 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
6208 mx->mx_cursor.mc_top = 0;
6209 mx->mx_cursor.mc_pg[0] = fp;
6210 mx->mx_cursor.mc_ki[0] = 0;
6211 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6212 mx->mx_db.md_flags = MDB_DUPFIXED;
6213 mx->mx_db.md_pad = fp->mp_pad;
6214 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6215 mx->mx_db.md_flags |= MDB_INTEGERKEY;
6218 DPRINTF("Sub-db %u for db %u root page %zu", mx->mx_cursor.mc_dbi, mc->mc_dbi,
6220 mx->mx_dbflag = DB_VALID | (F_ISSET(mc->mc_pg[mc->mc_top]->mp_flags, P_DIRTY) ?
6222 mx->mx_dbx.md_name.mv_data = NODEKEY(node);
6223 mx->mx_dbx.md_name.mv_size = node->mn_ksize;
6224 #if UINT_MAX < SIZE_MAX
6225 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
6226 #ifdef MISALIGNED_OK
6227 mx->mx_dbx.md_cmp = mdb_cmp_long;
6229 mx->mx_dbx.md_cmp = mdb_cmp_cint;
6234 /** Initialize a cursor for a given transaction and database. */
6236 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
6239 mc->mc_backup = NULL;
6242 mc->mc_db = &txn->mt_dbs[dbi];
6243 mc->mc_dbx = &txn->mt_dbxs[dbi];
6244 mc->mc_dbflag = &txn->mt_dbflags[dbi];
6249 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
6251 mc->mc_xcursor = mx;
6252 mdb_xcursor_init0(mc);
6254 mc->mc_xcursor = NULL;
6256 if (*mc->mc_dbflag & DB_STALE) {
6257 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
6262 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
6265 size_t size = sizeof(MDB_cursor);
6267 if (txn == NULL || ret == NULL || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
6270 /* Allow read access to the freelist */
6271 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
6274 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
6275 size += sizeof(MDB_xcursor);
6277 if ((mc = malloc(size)) != NULL) {
6278 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
6279 if (txn->mt_cursors) {
6280 mc->mc_next = txn->mt_cursors[dbi];
6281 txn->mt_cursors[dbi] = mc;
6282 mc->mc_flags |= C_UNTRACK;
6294 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
6296 if (txn == NULL || mc == NULL || mc->mc_dbi >= txn->mt_numdbs)
6299 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
6302 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
6306 /* Return the count of duplicate data items for the current key */
6308 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
6312 if (mc == NULL || countp == NULL)
6315 if (!(mc->mc_db->md_flags & MDB_DUPSORT))
6318 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6319 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6322 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
6325 *countp = mc->mc_xcursor->mx_db.md_entries;
6331 mdb_cursor_close(MDB_cursor *mc)
6333 if (mc && !mc->mc_backup) {
6334 /* remove from txn, if tracked */
6335 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
6336 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
6337 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
6339 *prev = mc->mc_next;
6346 mdb_cursor_txn(MDB_cursor *mc)
6348 if (!mc) return NULL;
6353 mdb_cursor_dbi(MDB_cursor *mc)
6359 /** Replace the key for a node with a new key.
6360 * @param[in] mc Cursor pointing to the node to operate on.
6361 * @param[in] key The new key to use.
6362 * @return 0 on success, non-zero on failure.
6365 mdb_update_key(MDB_cursor *mc, MDB_val *key)
6372 indx_t ptr, i, numkeys, indx;
6375 indx = mc->mc_ki[mc->mc_top];
6376 mp = mc->mc_pg[mc->mc_top];
6377 node = NODEPTR(mp, indx);
6378 ptr = mp->mp_ptrs[indx];
6382 char kbuf2[(MDB_MAXKEYSIZE*2+1)];
6383 k2.mv_data = NODEKEY(node);
6384 k2.mv_size = node->mn_ksize;
6385 DPRINTF("update key %u (ofs %u) [%s] to [%s] on page %zu",
6387 mdb_dkey(&k2, kbuf2),
6393 delta0 = delta = key->mv_size - node->mn_ksize;
6395 /* Must be 2-byte aligned. If new key is
6396 * shorter by 1, the shift will be skipped.
6398 delta += (delta & 1);
6400 if (delta > 0 && SIZELEFT(mp) < delta) {
6402 /* not enough space left, do a delete and split */
6403 DPRINTF("Not enough room, delta = %d, splitting...", delta);
6404 pgno = NODEPGNO(node);
6405 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
6406 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
6409 numkeys = NUMKEYS(mp);
6410 for (i = 0; i < numkeys; i++) {
6411 if (mp->mp_ptrs[i] <= ptr)
6412 mp->mp_ptrs[i] -= delta;
6415 base = (char *)mp + mp->mp_upper;
6416 len = ptr - mp->mp_upper + NODESIZE;
6417 memmove(base - delta, base, len);
6418 mp->mp_upper -= delta;
6420 node = NODEPTR(mp, indx);
6423 /* But even if no shift was needed, update ksize */
6425 node->mn_ksize = key->mv_size;
6428 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6434 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
6436 /** Move a node from csrc to cdst.
6439 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
6446 unsigned short flags;
6450 /* Mark src and dst as dirty. */
6451 if ((rc = mdb_page_touch(csrc)) ||
6452 (rc = mdb_page_touch(cdst)))
6455 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6456 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0); /* fake */
6457 key.mv_size = csrc->mc_db->md_pad;
6458 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
6460 data.mv_data = NULL;
6464 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
6465 assert(!((long)srcnode&1));
6466 srcpg = NODEPGNO(srcnode);
6467 flags = srcnode->mn_flags;
6468 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6469 unsigned int snum = csrc->mc_snum;
6471 /* must find the lowest key below src */
6472 mdb_page_search_lowest(csrc);
6473 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6474 key.mv_size = csrc->mc_db->md_pad;
6475 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6477 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6478 key.mv_size = NODEKSZ(s2);
6479 key.mv_data = NODEKEY(s2);
6481 csrc->mc_snum = snum--;
6482 csrc->mc_top = snum;
6484 key.mv_size = NODEKSZ(srcnode);
6485 key.mv_data = NODEKEY(srcnode);
6487 data.mv_size = NODEDSZ(srcnode);
6488 data.mv_data = NODEDATA(srcnode);
6490 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
6491 unsigned int snum = cdst->mc_snum;
6494 /* must find the lowest key below dst */
6495 mdb_page_search_lowest(cdst);
6496 if (IS_LEAF2(cdst->mc_pg[cdst->mc_top])) {
6497 bkey.mv_size = cdst->mc_db->md_pad;
6498 bkey.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, bkey.mv_size);
6500 s2 = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
6501 bkey.mv_size = NODEKSZ(s2);
6502 bkey.mv_data = NODEKEY(s2);
6504 cdst->mc_snum = snum--;
6505 cdst->mc_top = snum;
6506 mdb_cursor_copy(cdst, &mn);
6508 rc = mdb_update_key(&mn, &bkey);
6513 DPRINTF("moving %s node %u [%s] on page %zu to node %u on page %zu",
6514 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
6515 csrc->mc_ki[csrc->mc_top],
6517 csrc->mc_pg[csrc->mc_top]->mp_pgno,
6518 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno);
6520 /* Add the node to the destination page.
6522 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
6523 if (rc != MDB_SUCCESS)
6526 /* Delete the node from the source page.
6528 mdb_node_del(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
6531 /* Adjust other cursors pointing to mp */
6532 MDB_cursor *m2, *m3;
6533 MDB_dbi dbi = csrc->mc_dbi;
6534 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
6536 if (csrc->mc_flags & C_SUB)
6539 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6540 if (csrc->mc_flags & C_SUB)
6541 m3 = &m2->mc_xcursor->mx_cursor;
6544 if (m3 == csrc) continue;
6545 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
6546 csrc->mc_ki[csrc->mc_top]) {
6547 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
6548 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
6553 /* Update the parent separators.
6555 if (csrc->mc_ki[csrc->mc_top] == 0) {
6556 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
6557 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6558 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6560 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6561 key.mv_size = NODEKSZ(srcnode);
6562 key.mv_data = NODEKEY(srcnode);
6564 DPRINTF("update separator for source page %zu to [%s]",
6565 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key));
6566 mdb_cursor_copy(csrc, &mn);
6569 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
6572 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6574 indx_t ix = csrc->mc_ki[csrc->mc_top];
6575 nullkey.mv_size = 0;
6576 csrc->mc_ki[csrc->mc_top] = 0;
6577 rc = mdb_update_key(csrc, &nullkey);
6578 csrc->mc_ki[csrc->mc_top] = ix;
6579 assert(rc == MDB_SUCCESS);
6583 if (cdst->mc_ki[cdst->mc_top] == 0) {
6584 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
6585 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6586 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
6588 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
6589 key.mv_size = NODEKSZ(srcnode);
6590 key.mv_data = NODEKEY(srcnode);
6592 DPRINTF("update separator for destination page %zu to [%s]",
6593 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key));
6594 mdb_cursor_copy(cdst, &mn);
6597 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
6600 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
6602 indx_t ix = cdst->mc_ki[cdst->mc_top];
6603 nullkey.mv_size = 0;
6604 cdst->mc_ki[cdst->mc_top] = 0;
6605 rc = mdb_update_key(cdst, &nullkey);
6606 cdst->mc_ki[cdst->mc_top] = ix;
6607 assert(rc == MDB_SUCCESS);
6614 /** Merge one page into another.
6615 * The nodes from the page pointed to by \b csrc will
6616 * be copied to the page pointed to by \b cdst and then
6617 * the \b csrc page will be freed.
6618 * @param[in] csrc Cursor pointing to the source page.
6619 * @param[in] cdst Cursor pointing to the destination page.
6622 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
6630 DPRINTF("merging page %zu into %zu", csrc->mc_pg[csrc->mc_top]->mp_pgno,
6631 cdst->mc_pg[cdst->mc_top]->mp_pgno);
6633 assert(csrc->mc_snum > 1); /* can't merge root page */
6634 assert(cdst->mc_snum > 1);
6636 /* Mark dst as dirty. */
6637 if ((rc = mdb_page_touch(cdst)))
6640 /* Move all nodes from src to dst.
6642 j = nkeys = NUMKEYS(cdst->mc_pg[cdst->mc_top]);
6643 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6644 key.mv_size = csrc->mc_db->md_pad;
6645 key.mv_data = METADATA(csrc->mc_pg[csrc->mc_top]);
6646 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
6647 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
6648 if (rc != MDB_SUCCESS)
6650 key.mv_data = (char *)key.mv_data + key.mv_size;
6653 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
6654 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], i);
6655 if (i == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6656 unsigned int snum = csrc->mc_snum;
6658 /* must find the lowest key below src */
6659 mdb_page_search_lowest(csrc);
6660 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6661 key.mv_size = csrc->mc_db->md_pad;
6662 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6664 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6665 key.mv_size = NODEKSZ(s2);
6666 key.mv_data = NODEKEY(s2);
6668 csrc->mc_snum = snum--;
6669 csrc->mc_top = snum;
6671 key.mv_size = srcnode->mn_ksize;
6672 key.mv_data = NODEKEY(srcnode);
6675 data.mv_size = NODEDSZ(srcnode);
6676 data.mv_data = NODEDATA(srcnode);
6677 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
6678 if (rc != MDB_SUCCESS)
6683 DPRINTF("dst page %zu now has %u keys (%.1f%% filled)",
6684 cdst->mc_pg[cdst->mc_top]->mp_pgno, NUMKEYS(cdst->mc_pg[cdst->mc_top]), (float)PAGEFILL(cdst->mc_txn->mt_env, cdst->mc_pg[cdst->mc_top]) / 10);
6686 /* Unlink the src page from parent and add to free list.
6688 mdb_node_del(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1], 0);
6689 if (csrc->mc_ki[csrc->mc_top-1] == 0) {
6692 rc = mdb_update_key(csrc, &key);
6698 rc = mdb_midl_append(&csrc->mc_txn->mt_free_pgs,
6699 csrc->mc_pg[csrc->mc_top]->mp_pgno);
6702 if (IS_LEAF(csrc->mc_pg[csrc->mc_top]))
6703 csrc->mc_db->md_leaf_pages--;
6705 csrc->mc_db->md_branch_pages--;
6707 /* Adjust other cursors pointing to mp */
6708 MDB_cursor *m2, *m3;
6709 MDB_dbi dbi = csrc->mc_dbi;
6710 MDB_page *mp = cdst->mc_pg[cdst->mc_top];
6712 if (csrc->mc_flags & C_SUB)
6715 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6716 if (csrc->mc_flags & C_SUB)
6717 m3 = &m2->mc_xcursor->mx_cursor;
6720 if (m3 == csrc) continue;
6721 if (m3->mc_snum < csrc->mc_snum) continue;
6722 if (m3->mc_pg[csrc->mc_top] == csrc->mc_pg[csrc->mc_top]) {
6723 m3->mc_pg[csrc->mc_top] = mp;
6724 m3->mc_ki[csrc->mc_top] += nkeys;
6728 mdb_cursor_pop(csrc);
6730 return mdb_rebalance(csrc);
6733 /** Copy the contents of a cursor.
6734 * @param[in] csrc The cursor to copy from.
6735 * @param[out] cdst The cursor to copy to.
6738 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
6742 cdst->mc_txn = csrc->mc_txn;
6743 cdst->mc_dbi = csrc->mc_dbi;
6744 cdst->mc_db = csrc->mc_db;
6745 cdst->mc_dbx = csrc->mc_dbx;
6746 cdst->mc_snum = csrc->mc_snum;
6747 cdst->mc_top = csrc->mc_top;
6748 cdst->mc_flags = csrc->mc_flags;
6750 for (i=0; i<csrc->mc_snum; i++) {
6751 cdst->mc_pg[i] = csrc->mc_pg[i];
6752 cdst->mc_ki[i] = csrc->mc_ki[i];
6756 /** Rebalance the tree after a delete operation.
6757 * @param[in] mc Cursor pointing to the page where rebalancing
6759 * @return 0 on success, non-zero on failure.
6762 mdb_rebalance(MDB_cursor *mc)
6766 unsigned int ptop, minkeys;
6769 minkeys = 1 + (IS_BRANCH(mc->mc_pg[mc->mc_top]));
6773 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
6774 DPRINTF("rebalancing %s page %zu (has %u keys, %.1f%% full)",
6775 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
6776 pgno, NUMKEYS(mc->mc_pg[mc->mc_top]), (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10);
6780 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD &&
6781 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
6784 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
6785 DPRINTF("no need to rebalance page %zu, above fill threshold",
6791 if (mc->mc_snum < 2) {
6792 MDB_page *mp = mc->mc_pg[0];
6794 DPUTS("Can't rebalance a subpage, ignoring");
6797 if (NUMKEYS(mp) == 0) {
6798 DPUTS("tree is completely empty");
6799 mc->mc_db->md_root = P_INVALID;
6800 mc->mc_db->md_depth = 0;
6801 mc->mc_db->md_leaf_pages = 0;
6802 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
6805 /* Adjust cursors pointing to mp */
6809 MDB_cursor *m2, *m3;
6810 MDB_dbi dbi = mc->mc_dbi;
6812 if (mc->mc_flags & C_SUB)
6815 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6816 if (mc->mc_flags & C_SUB)
6817 m3 = &m2->mc_xcursor->mx_cursor;
6820 if (m3->mc_snum < mc->mc_snum) continue;
6821 if (m3->mc_pg[0] == mp) {
6827 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
6828 DPUTS("collapsing root page!");
6829 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
6832 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
6833 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
6836 mc->mc_db->md_depth--;
6837 mc->mc_db->md_branch_pages--;
6838 mc->mc_ki[0] = mc->mc_ki[1];
6840 /* Adjust other cursors pointing to mp */
6841 MDB_cursor *m2, *m3;
6842 MDB_dbi dbi = mc->mc_dbi;
6844 if (mc->mc_flags & C_SUB)
6847 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6848 if (mc->mc_flags & C_SUB)
6849 m3 = &m2->mc_xcursor->mx_cursor;
6852 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
6853 if (m3->mc_pg[0] == mp) {
6854 m3->mc_pg[0] = mc->mc_pg[0];
6857 m3->mc_ki[0] = m3->mc_ki[1];
6862 DPUTS("root page doesn't need rebalancing");
6866 /* The parent (branch page) must have at least 2 pointers,
6867 * otherwise the tree is invalid.
6869 ptop = mc->mc_top-1;
6870 assert(NUMKEYS(mc->mc_pg[ptop]) > 1);
6872 /* Leaf page fill factor is below the threshold.
6873 * Try to move keys from left or right neighbor, or
6874 * merge with a neighbor page.
6879 mdb_cursor_copy(mc, &mn);
6880 mn.mc_xcursor = NULL;
6882 if (mc->mc_ki[ptop] == 0) {
6883 /* We're the leftmost leaf in our parent.
6885 DPUTS("reading right neighbor");
6887 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
6888 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
6891 mn.mc_ki[mn.mc_top] = 0;
6892 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
6894 /* There is at least one neighbor to the left.
6896 DPUTS("reading left neighbor");
6898 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
6899 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
6902 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
6903 mc->mc_ki[mc->mc_top] = 0;
6906 DPRINTF("found neighbor page %zu (%u keys, %.1f%% full)",
6907 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]), (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10);
6909 /* If the neighbor page is above threshold and has enough keys,
6910 * move one key from it. Otherwise we should try to merge them.
6911 * (A branch page must never have less than 2 keys.)
6913 minkeys = 1 + (IS_BRANCH(mn.mc_pg[mn.mc_top]));
6914 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys)
6915 return mdb_node_move(&mn, mc);
6917 if (mc->mc_ki[ptop] == 0)
6918 rc = mdb_page_merge(&mn, mc);
6920 rc = mdb_page_merge(mc, &mn);
6921 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
6926 /** Complete a delete operation started by #mdb_cursor_del(). */
6928 mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf)
6934 mp = mc->mc_pg[mc->mc_top];
6935 ki = mc->mc_ki[mc->mc_top];
6937 /* add overflow pages to free list */
6938 if (!IS_LEAF2(mp) && F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6942 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
6943 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) ||
6944 (rc = mdb_ovpage_free(mc, omp)))
6947 mdb_node_del(mp, ki, mc->mc_db->md_pad);
6948 mc->mc_db->md_entries--;
6949 rc = mdb_rebalance(mc);
6950 if (rc != MDB_SUCCESS)
6951 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6952 /* if mc points past last node in page, invalidate */
6953 else if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
6954 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
6957 /* Adjust other cursors pointing to mp */
6960 MDB_dbi dbi = mc->mc_dbi;
6962 mp = mc->mc_pg[mc->mc_top];
6963 nkeys = NUMKEYS(mp);
6964 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6967 if (!(m2->mc_flags & C_INITIALIZED))
6969 if (m2->mc_pg[mc->mc_top] == mp) {
6970 if (m2->mc_ki[mc->mc_top] > ki)
6971 m2->mc_ki[mc->mc_top]--;
6972 if (m2->mc_ki[mc->mc_top] >= nkeys)
6973 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
6982 mdb_del(MDB_txn *txn, MDB_dbi dbi,
6983 MDB_val *key, MDB_val *data)
6988 MDB_val rdata, *xdata;
6992 assert(key != NULL);
6994 DPRINTF("====> delete db %u key [%s]", dbi, DKEY(key));
6996 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
6999 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
7003 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
7007 mdb_cursor_init(&mc, txn, dbi, &mx);
7018 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
7020 /* let mdb_page_split know about this cursor if needed:
7021 * delete will trigger a rebalance; if it needs to move
7022 * a node from one page to another, it will have to
7023 * update the parent's separator key(s). If the new sepkey
7024 * is larger than the current one, the parent page may
7025 * run out of space, triggering a split. We need this
7026 * cursor to be consistent until the end of the rebalance.
7028 mc.mc_flags |= C_UNTRACK;
7029 mc.mc_next = txn->mt_cursors[dbi];
7030 txn->mt_cursors[dbi] = &mc;
7031 rc = mdb_cursor_del(&mc, data ? 0 : MDB_NODUPDATA);
7032 txn->mt_cursors[dbi] = mc.mc_next;
7037 /** Split a page and insert a new node.
7038 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
7039 * The cursor will be updated to point to the actual page and index where
7040 * the node got inserted after the split.
7041 * @param[in] newkey The key for the newly inserted node.
7042 * @param[in] newdata The data for the newly inserted node.
7043 * @param[in] newpgno The page number, if the new node is a branch node.
7044 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
7045 * @return 0 on success, non-zero on failure.
7048 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
7049 unsigned int nflags)
7052 int rc = MDB_SUCCESS, ins_new = 0, new_root = 0, newpos = 1, did_split = 0;
7055 unsigned int i, j, split_indx, nkeys, pmax;
7057 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
7059 MDB_page *mp, *rp, *pp;
7064 mp = mc->mc_pg[mc->mc_top];
7065 newindx = mc->mc_ki[mc->mc_top];
7067 DPRINTF("-----> splitting %s page %zu and adding [%s] at index %i",
7068 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
7069 DKEY(newkey), mc->mc_ki[mc->mc_top]);
7071 /* Create a right sibling. */
7072 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
7074 DPRINTF("new right sibling: page %zu", rp->mp_pgno);
7076 if (mc->mc_snum < 2) {
7077 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
7079 /* shift current top to make room for new parent */
7080 mc->mc_pg[1] = mc->mc_pg[0];
7081 mc->mc_ki[1] = mc->mc_ki[0];
7084 mc->mc_db->md_root = pp->mp_pgno;
7085 DPRINTF("root split! new root = %zu", pp->mp_pgno);
7086 mc->mc_db->md_depth++;
7089 /* Add left (implicit) pointer. */
7090 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
7091 /* undo the pre-push */
7092 mc->mc_pg[0] = mc->mc_pg[1];
7093 mc->mc_ki[0] = mc->mc_ki[1];
7094 mc->mc_db->md_root = mp->mp_pgno;
7095 mc->mc_db->md_depth--;
7102 ptop = mc->mc_top-1;
7103 DPRINTF("parent branch page is %zu", mc->mc_pg[ptop]->mp_pgno);
7106 mc->mc_flags |= C_SPLITTING;
7107 mdb_cursor_copy(mc, &mn);
7108 mn.mc_pg[mn.mc_top] = rp;
7109 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
7111 if (nflags & MDB_APPEND) {
7112 mn.mc_ki[mn.mc_top] = 0;
7114 split_indx = newindx;
7119 nkeys = NUMKEYS(mp);
7120 split_indx = nkeys / 2;
7121 if (newindx < split_indx)
7127 unsigned int lsize, rsize, ksize;
7128 /* Move half of the keys to the right sibling */
7130 x = mc->mc_ki[mc->mc_top] - split_indx;
7131 ksize = mc->mc_db->md_pad;
7132 split = LEAF2KEY(mp, split_indx, ksize);
7133 rsize = (nkeys - split_indx) * ksize;
7134 lsize = (nkeys - split_indx) * sizeof(indx_t);
7135 mp->mp_lower -= lsize;
7136 rp->mp_lower += lsize;
7137 mp->mp_upper += rsize - lsize;
7138 rp->mp_upper -= rsize - lsize;
7139 sepkey.mv_size = ksize;
7140 if (newindx == split_indx) {
7141 sepkey.mv_data = newkey->mv_data;
7143 sepkey.mv_data = split;
7146 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
7147 memcpy(rp->mp_ptrs, split, rsize);
7148 sepkey.mv_data = rp->mp_ptrs;
7149 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
7150 memcpy(ins, newkey->mv_data, ksize);
7151 mp->mp_lower += sizeof(indx_t);
7152 mp->mp_upper -= ksize - sizeof(indx_t);
7155 memcpy(rp->mp_ptrs, split, x * ksize);
7156 ins = LEAF2KEY(rp, x, ksize);
7157 memcpy(ins, newkey->mv_data, ksize);
7158 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
7159 rp->mp_lower += sizeof(indx_t);
7160 rp->mp_upper -= ksize - sizeof(indx_t);
7161 mc->mc_ki[mc->mc_top] = x;
7162 mc->mc_pg[mc->mc_top] = rp;
7167 /* For leaf pages, check the split point based on what
7168 * fits where, since otherwise mdb_node_add can fail.
7170 * This check is only needed when the data items are
7171 * relatively large, such that being off by one will
7172 * make the difference between success or failure.
7174 * It's also relevant if a page happens to be laid out
7175 * such that one half of its nodes are all "small" and
7176 * the other half of its nodes are "large." If the new
7177 * item is also "large" and falls on the half with
7178 * "large" nodes, it also may not fit.
7181 unsigned int psize, nsize;
7182 /* Maximum free space in an empty page */
7183 pmax = mc->mc_txn->mt_env->me_psize - PAGEHDRSZ;
7184 nsize = mdb_leaf_size(mc->mc_txn->mt_env, newkey, newdata);
7185 if ((nkeys < 20) || (nsize > pmax/16)) {
7186 if (newindx <= split_indx) {
7189 for (i=0; i<split_indx; i++) {
7190 node = NODEPTR(mp, i);
7191 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
7192 if (F_ISSET(node->mn_flags, F_BIGDATA))
7193 psize += sizeof(pgno_t);
7195 psize += NODEDSZ(node);
7199 split_indx = newindx;
7210 for (i=nkeys-1; i>=split_indx; i--) {
7211 node = NODEPTR(mp, i);
7212 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
7213 if (F_ISSET(node->mn_flags, F_BIGDATA))
7214 psize += sizeof(pgno_t);
7216 psize += NODEDSZ(node);
7220 split_indx = newindx;
7231 /* First find the separating key between the split pages.
7232 * The case where newindx == split_indx is ambiguous; the
7233 * new item could go to the new page or stay on the original
7234 * page. If newpos == 1 it goes to the new page.
7236 if (newindx == split_indx && newpos) {
7237 sepkey.mv_size = newkey->mv_size;
7238 sepkey.mv_data = newkey->mv_data;
7240 node = NODEPTR(mp, split_indx);
7241 sepkey.mv_size = node->mn_ksize;
7242 sepkey.mv_data = NODEKEY(node);
7246 DPRINTF("separator is [%s]", DKEY(&sepkey));
7248 /* Copy separator key to the parent.
7250 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(mc->mc_txn->mt_env, &sepkey)) {
7254 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
7257 if (mn.mc_snum == mc->mc_snum) {
7258 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
7259 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
7260 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
7261 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
7266 /* Right page might now have changed parent.
7267 * Check if left page also changed parent.
7269 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
7270 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
7271 for (i=0; i<ptop; i++) {
7272 mc->mc_pg[i] = mn.mc_pg[i];
7273 mc->mc_ki[i] = mn.mc_ki[i];
7275 mc->mc_pg[ptop] = mn.mc_pg[ptop];
7276 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
7280 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
7283 mc->mc_flags ^= C_SPLITTING;
7284 if (rc != MDB_SUCCESS) {
7287 if (nflags & MDB_APPEND) {
7288 mc->mc_pg[mc->mc_top] = rp;
7289 mc->mc_ki[mc->mc_top] = 0;
7290 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
7293 for (i=0; i<mc->mc_top; i++)
7294 mc->mc_ki[i] = mn.mc_ki[i];
7301 /* Move half of the keys to the right sibling. */
7303 /* grab a page to hold a temporary copy */
7304 copy = mdb_page_malloc(mc->mc_txn, 1);
7308 copy->mp_pgno = mp->mp_pgno;
7309 copy->mp_flags = mp->mp_flags;
7310 copy->mp_lower = PAGEHDRSZ;
7311 copy->mp_upper = mc->mc_txn->mt_env->me_psize;
7312 mc->mc_pg[mc->mc_top] = copy;
7313 for (i = j = 0; i <= nkeys; j++) {
7314 if (i == split_indx) {
7315 /* Insert in right sibling. */
7316 /* Reset insert index for right sibling. */
7317 if (i != newindx || (newpos ^ ins_new)) {
7319 mc->mc_pg[mc->mc_top] = rp;
7323 if (i == newindx && !ins_new) {
7324 /* Insert the original entry that caused the split. */
7325 rkey.mv_data = newkey->mv_data;
7326 rkey.mv_size = newkey->mv_size;
7335 /* Update index for the new key. */
7336 mc->mc_ki[mc->mc_top] = j;
7337 } else if (i == nkeys) {
7340 node = NODEPTR(mp, i);
7341 rkey.mv_data = NODEKEY(node);
7342 rkey.mv_size = node->mn_ksize;
7344 xdata.mv_data = NODEDATA(node);
7345 xdata.mv_size = NODEDSZ(node);
7348 pgno = NODEPGNO(node);
7349 flags = node->mn_flags;
7354 if (!IS_LEAF(mp) && j == 0) {
7355 /* First branch index doesn't need key data. */
7359 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
7363 nkeys = NUMKEYS(copy);
7364 for (i=0; i<nkeys; i++)
7365 mp->mp_ptrs[i] = copy->mp_ptrs[i];
7366 mp->mp_lower = copy->mp_lower;
7367 mp->mp_upper = copy->mp_upper;
7368 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
7369 mc->mc_txn->mt_env->me_psize - copy->mp_upper);
7371 /* reset back to original page */
7372 if (newindx < split_indx || (!newpos && newindx == split_indx)) {
7373 mc->mc_pg[mc->mc_top] = mp;
7374 if (nflags & MDB_RESERVE) {
7375 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
7376 if (!(node->mn_flags & F_BIGDATA))
7377 newdata->mv_data = NODEDATA(node);
7381 /* Make sure mc_ki is still valid.
7383 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
7384 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
7385 for (i=0; i<ptop; i++) {
7386 mc->mc_pg[i] = mn.mc_pg[i];
7387 mc->mc_ki[i] = mn.mc_ki[i];
7389 mc->mc_pg[ptop] = mn.mc_pg[ptop];
7390 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
7394 /* return tmp page to freelist */
7395 mdb_page_free(mc->mc_txn->mt_env, copy);
7398 /* Adjust other cursors pointing to mp */
7399 MDB_cursor *m2, *m3;
7400 MDB_dbi dbi = mc->mc_dbi;
7401 int fixup = NUMKEYS(mp);
7403 if (mc->mc_flags & C_SUB)
7406 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7407 if (mc->mc_flags & C_SUB)
7408 m3 = &m2->mc_xcursor->mx_cursor;
7413 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
7415 if (m3->mc_flags & C_SPLITTING)
7420 for (k=m3->mc_top; k>=0; k--) {
7421 m3->mc_ki[k+1] = m3->mc_ki[k];
7422 m3->mc_pg[k+1] = m3->mc_pg[k];
7424 if (m3->mc_ki[0] >= split_indx) {
7429 m3->mc_pg[0] = mc->mc_pg[0];
7433 if (m3->mc_pg[mc->mc_top] == mp) {
7434 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
7435 m3->mc_ki[mc->mc_top]++;
7436 if (m3->mc_ki[mc->mc_top] >= fixup) {
7437 m3->mc_pg[mc->mc_top] = rp;
7438 m3->mc_ki[mc->mc_top] -= fixup;
7439 m3->mc_ki[ptop] = mn.mc_ki[ptop];
7441 } else if (!did_split && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
7442 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
7451 mdb_put(MDB_txn *txn, MDB_dbi dbi,
7452 MDB_val *key, MDB_val *data, unsigned int flags)
7457 assert(key != NULL);
7458 assert(data != NULL);
7460 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7463 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
7467 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
7471 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP)) != flags)
7474 mdb_cursor_init(&mc, txn, dbi, &mx);
7475 return mdb_cursor_put(&mc, key, data, flags);
7479 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
7481 if ((flag & CHANGEABLE) != flag)
7484 env->me_flags |= flag;
7486 env->me_flags &= ~flag;
7491 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
7496 *arg = env->me_flags;
7501 mdb_env_get_path(MDB_env *env, const char **arg)
7506 *arg = env->me_path;
7510 /** Common code for #mdb_stat() and #mdb_env_stat().
7511 * @param[in] env the environment to operate in.
7512 * @param[in] db the #MDB_db record containing the stats to return.
7513 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
7514 * @return 0, this function always succeeds.
7517 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
7519 arg->ms_psize = env->me_psize;
7520 arg->ms_depth = db->md_depth;
7521 arg->ms_branch_pages = db->md_branch_pages;
7522 arg->ms_leaf_pages = db->md_leaf_pages;
7523 arg->ms_overflow_pages = db->md_overflow_pages;
7524 arg->ms_entries = db->md_entries;
7529 mdb_env_stat(MDB_env *env, MDB_stat *arg)
7533 if (env == NULL || arg == NULL)
7536 toggle = mdb_env_pick_meta(env);
7538 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
7542 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
7546 if (env == NULL || arg == NULL)
7549 toggle = mdb_env_pick_meta(env);
7550 arg->me_mapaddr = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : 0;
7551 arg->me_mapsize = env->me_mapsize;
7552 arg->me_maxreaders = env->me_maxreaders;
7553 arg->me_numreaders = env->me_numreaders;
7554 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
7555 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
7559 /** Set the default comparison functions for a database.
7560 * Called immediately after a database is opened to set the defaults.
7561 * The user can then override them with #mdb_set_compare() or
7562 * #mdb_set_dupsort().
7563 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
7564 * @param[in] dbi A database handle returned by #mdb_dbi_open()
7567 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
7569 uint16_t f = txn->mt_dbs[dbi].md_flags;
7571 txn->mt_dbxs[dbi].md_cmp =
7572 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
7573 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
7575 txn->mt_dbxs[dbi].md_dcmp =
7576 !(f & MDB_DUPSORT) ? 0 :
7577 ((f & MDB_INTEGERDUP)
7578 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
7579 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
7582 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
7587 int rc, dbflag, exact;
7588 unsigned int unused = 0;
7591 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
7592 mdb_default_cmp(txn, FREE_DBI);
7595 if ((flags & VALID_FLAGS) != flags)
7601 if (flags & PERSISTENT_FLAGS) {
7602 uint16_t f2 = flags & PERSISTENT_FLAGS;
7603 /* make sure flag changes get committed */
7604 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
7605 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
7606 txn->mt_flags |= MDB_TXN_DIRTY;
7609 mdb_default_cmp(txn, MAIN_DBI);
7613 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
7614 mdb_default_cmp(txn, MAIN_DBI);
7617 /* Is the DB already open? */
7619 for (i=2; i<txn->mt_numdbs; i++) {
7620 if (!txn->mt_dbxs[i].md_name.mv_size) {
7621 /* Remember this free slot */
7622 if (!unused) unused = i;
7625 if (len == txn->mt_dbxs[i].md_name.mv_size &&
7626 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
7632 /* If no free slot and max hit, fail */
7633 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
7634 return MDB_DBS_FULL;
7636 /* Cannot mix named databases with some mainDB flags */
7637 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
7638 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
7640 /* Find the DB info */
7641 dbflag = DB_NEW|DB_VALID;
7644 key.mv_data = (void *)name;
7645 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
7646 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
7647 if (rc == MDB_SUCCESS) {
7648 /* make sure this is actually a DB */
7649 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
7650 if (!(node->mn_flags & F_SUBDATA))
7652 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
7653 /* Create if requested */
7655 data.mv_size = sizeof(MDB_db);
7656 data.mv_data = &dummy;
7657 memset(&dummy, 0, sizeof(dummy));
7658 dummy.md_root = P_INVALID;
7659 dummy.md_flags = flags & PERSISTENT_FLAGS;
7660 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
7664 /* OK, got info, add to table */
7665 if (rc == MDB_SUCCESS) {
7666 unsigned int slot = unused ? unused : txn->mt_numdbs;
7667 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
7668 txn->mt_dbxs[slot].md_name.mv_size = len;
7669 txn->mt_dbxs[slot].md_rel = NULL;
7670 txn->mt_dbflags[slot] = dbflag;
7671 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
7673 txn->mt_env->me_dbflags[slot] = txn->mt_dbs[slot].md_flags;
7674 mdb_default_cmp(txn, slot);
7683 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
7685 if (txn == NULL || arg == NULL || dbi >= txn->mt_numdbs)
7688 if (txn->mt_dbflags[dbi] & DB_STALE) {
7691 /* Stale, must read the DB's root. cursor_init does it for us. */
7692 mdb_cursor_init(&mc, txn, dbi, &mx);
7694 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
7697 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
7700 if (dbi <= MAIN_DBI || dbi >= env->me_maxdbs)
7702 ptr = env->me_dbxs[dbi].md_name.mv_data;
7703 env->me_dbxs[dbi].md_name.mv_data = NULL;
7704 env->me_dbxs[dbi].md_name.mv_size = 0;
7705 env->me_dbflags[dbi] = 0;
7709 /** Add all the DB's pages to the free list.
7710 * @param[in] mc Cursor on the DB to free.
7711 * @param[in] subs non-Zero to check for sub-DBs in this DB.
7712 * @return 0 on success, non-zero on failure.
7715 mdb_drop0(MDB_cursor *mc, int subs)
7719 rc = mdb_page_search(mc, NULL, 0);
7720 if (rc == MDB_SUCCESS) {
7721 MDB_txn *txn = mc->mc_txn;
7726 /* LEAF2 pages have no nodes, cannot have sub-DBs */
7727 if (IS_LEAF2(mc->mc_pg[mc->mc_top]))
7730 mdb_cursor_copy(mc, &mx);
7731 while (mc->mc_snum > 0) {
7732 MDB_page *mp = mc->mc_pg[mc->mc_top];
7733 unsigned n = NUMKEYS(mp);
7735 for (i=0; i<n; i++) {
7736 ni = NODEPTR(mp, i);
7737 if (ni->mn_flags & F_BIGDATA) {
7740 memcpy(&pg, NODEDATA(ni), sizeof(pg));
7741 rc = mdb_page_get(txn, pg, &omp, NULL);
7744 assert(IS_OVERFLOW(omp));
7745 rc = mdb_midl_append_range(&txn->mt_free_pgs,
7749 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
7750 mdb_xcursor_init1(mc, ni);
7751 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
7757 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
7759 for (i=0; i<n; i++) {
7761 ni = NODEPTR(mp, i);
7764 mdb_midl_xappend(txn->mt_free_pgs, pg);
7769 mc->mc_ki[mc->mc_top] = i;
7770 rc = mdb_cursor_sibling(mc, 1);
7772 /* no more siblings, go back to beginning
7773 * of previous level.
7777 for (i=1; i<mc->mc_snum; i++) {
7779 mc->mc_pg[i] = mx.mc_pg[i];
7784 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
7785 } else if (rc == MDB_NOTFOUND) {
7791 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
7793 MDB_cursor *mc, *m2;
7796 if (!txn || !dbi || dbi >= txn->mt_numdbs || (unsigned)del > 1 || !(txn->mt_dbflags[dbi] & DB_VALID))
7799 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
7802 rc = mdb_cursor_open(txn, dbi, &mc);
7806 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
7807 /* Invalidate the dropped DB's cursors */
7808 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
7809 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
7813 /* Can't delete the main DB */
7814 if (del && dbi > MAIN_DBI) {
7815 rc = mdb_del(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL);
7817 txn->mt_dbflags[dbi] = DB_STALE;
7818 mdb_dbi_close(txn->mt_env, dbi);
7821 /* reset the DB record, mark it dirty */
7822 txn->mt_dbflags[dbi] |= DB_DIRTY;
7823 txn->mt_dbs[dbi].md_depth = 0;
7824 txn->mt_dbs[dbi].md_branch_pages = 0;
7825 txn->mt_dbs[dbi].md_leaf_pages = 0;
7826 txn->mt_dbs[dbi].md_overflow_pages = 0;
7827 txn->mt_dbs[dbi].md_entries = 0;
7828 txn->mt_dbs[dbi].md_root = P_INVALID;
7830 txn->mt_flags |= MDB_TXN_DIRTY;
7833 mdb_cursor_close(mc);
7837 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
7839 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7842 txn->mt_dbxs[dbi].md_cmp = cmp;
7846 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
7848 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7851 txn->mt_dbxs[dbi].md_dcmp = cmp;
7855 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
7857 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7860 txn->mt_dbxs[dbi].md_rel = rel;
7864 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
7866 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7869 txn->mt_dbxs[dbi].md_relctx = ctx;